Knowledge Graph Theory and Structural Parsing Lei Zhang

2002 Ph.D. thesis University of Twente
Twente University Press

Also available in print:


Publisher: Twente University Press, P.O. Box 217, 7500 AE Enschede, The Netherlands, Print: Océ Facility Services, Enschede c L. Zhang, Enschede 2002

No part of this work may be reproduced by print, photocopy or any other means without the permission in writing from the publisher. ISBN 9036518350



ter verkrijging van de graad van doctor aan de Universiteit Twente, op gezag van de rector magnificus, prof. dr. F. A. van Vught, volgens besluit van het College voor Promoties in het openbaar te verdedigen op woensdag 20 november 2002 te 16:45 uur door Lei Zhang geboren op 10 maart 1964 te Xi’an, China

C. dr.Die proefschrift is goedgekeurd door de promotoren prof. Li . Hoede en prof. dr. X.

I am indebted to my parents . namely Xiaodong Liu and Shenggui Zhang for making my time in Twente so enjoyable. who gave me the opportunity to visit to Twente University in the Netherlands. very special thanks go to my family and all my friends. All those emails. not only from a scientific point of view. Working with other people speeds up the research process. He encouraged and helped me in everything I did for my thesis.Acknowledgements In the first place I express my gratitude to Prof. Xueliang Li as my cosupervisor. He has been of great help in introducing interesting areas of knowledge graph theory to me. students at Twente University. by his patience.D. Finally. This thesis benefited much from his detailed criticism and advice. I consider myself very lucky to have had Prof. It has been most pleasant to work with him. discussions and talks are unforgettable and were always a great motivation. my mother. Especially I like to thank Dr. my husband and my two sisters. He flooded my email box to help me continue my research while I was in China and we could not work face to face temporarily. Carla and Dini for all their help. Also I like to thank the other Ph. In particular I thank my father. I have learned a lot from him. understanding and caring. I like to thank all my colleagues for giving me a pleasant atmosphere to work. Broersma. Cornelis Hoede for inviting me to work under his inspiring supervision. With him I have had many extremely interesting and stimulating discussions.

It would have been impossible to have written this thesis without their unconditional support. I would like to thank my son for his sensibilities and love.VI Acknowledgements for their great support and encouragement. Enschede Lei Zhang . Most especially. October 2002.

but then started a rather large program of his own that led to the development of the medical expert system MEDES. In first instance only three types of relationships were distinguished. who. The other theory started in 1982 and was called a theory of knowledge graphs. The types of the arcs express relationships between concepts. who wanted to extract knowledge from medical and sociological texts. 1991] investigated robustness and consistency of extracted knowledge graphs.Preface This thesis makes a contribution to a theory of knowledge representation by means of graphs. that included a so-called path algebra. Their first PhD-student was Bakker [Bakker. One was started by Sowa. of which the causal relationship was the most important one. The theory belongs to the broad spectrum of semantic networks. in his thesis. In a third thesis Smit [Smit. by de Vries [de Vries. The knowledge graphs in that system had 18 types of . deals mainly with the problem of extracting causal relationships from a text. The second thesis. developed an information system. 1987] participated as well. 1987]. in order to obtain expert systems. the so-called total graph form in the terminology of Harary [Harary. 1989]. Concepts are represented by labeled vertices connected by labeled arcs. like AGENT or INSTRUMENT. to obtain implied relationships. In the eighties two related theories developed. 1972]. It was developed by Hoede and Stokman. who published a book on conceptual graphs in 1984. that are also represented by vertices. In the beginning of the project de Vries Robbé [de Vries Robbé.

This was one of the reasons that after the focus had been on the structuring of knowledge. This work can be seen as an extension of the work of Peirce [Peirce. representing both structured impressions from the “outer” world and his “inner” world. Now the problem of the ontology of knowledge graphs came forward. The increase in the number of types of relationships. The construction of word graphs was investigated by Hoede and students of the University of Twente. In principle every word should have a word graph. is due to the fact that many sentences of a text have to be left unprocessed if one considers only three types of relationships. 1993] was titled “Logic and Knowledge Graphs: one of two kinds” and deals with the problem of representing logical systems in terms of knowledge graphs. would study Chinese by means of knowledge graphs.VIII Preface relationships. 1885] on existential graphs. In fact. Arbitrary linguistic sentences should be representable by knowledge graphs. The interesting point here is that English and Chinese are significantly different as languages. By coincidence Li [Li. Liu and Zhang. leading to the first three theses. Back in China he proposed to start a joint project in which two students. a sentence should be represented by a sentence graph. Liu wrote a thesis that focuses on these specific . that also stands at the basis of Sowa’s theory of conceptual graphs. 1991] wrote a thesis in Twente on the purely mathematical subject of “Transformation Graphs”. man is considered to have in mind a mind graph. The thesis of Willems [Willems. the knowledge graph project was continued. This led to considerable extension of the number of types of relationships and to the introduction of so-called frames. then also specific features of Chinese should be representable within the theory. be considered to be too valuable to delete. 1993] was titled “Chemistry of Language” and started off the rather ambitious project of representing language by knowledge graphs. very much in line with what can be seen in semantic networks. If the paradigm that language can be expressed by knowledge graphs is to be defended. Several hundreds of the words frequently used in English were represented within the knowledge graph formalism. The thesis of van den Berg [Berg. however. focusing on the representation of knowledge in general. The information in these sentences may. The lexicon of word graphs is a prerequisite for any further investigation of language by means of graphs. a huge knowledge graph.



features as well as on other problems, like the extraction of causal relationships, which was done before, for English, by de Vries [de Vries, 1989]. In this thesis the focus is on the extraction of sentence graphs, both for English and for Chinese sentences. A prerequisite for this is a lexicon of word graphs. Three papers were dedicated to this. Hoede and Li [Hoede & Li, 1996] wrote a paper on a first set of words; verbs, nouns and prepositions. Hoede and Liu [Hoede & Liu, 1998] wrote a paper on a second set of words; adverbs, adjectives and Chinese classifiers or quantity words. Hoede and Zhang [Hoede & Zhang, 2001a] wrote a paper on a third set of words; the logic words, which is part of this thesis, see Chapter 4. In all three papers both Chinese and English words are considered. The contents of the first two papers are summarized in appendixes of this thesis. Mapping a sentence on a sentence graph, which is called structural parsing, constitutes the main theme of this thesis. A paper by Hoede and Zhang [Hoede & Zhang, 2001b] contains a shortened version of Chapter 5. Concepts involved in this thesis are e.g. semantic word graph, and syntactic word graph, utterance path and chunk. Utterance paths, studied in Chapter 6 and partial structural parsing, involving chunks and studied in Chapter 7, can be seen as first extensions of the developed theory in theoretical direction respectively applied direction. The main results and conclusions coming forward from our research are the following. i) Although there are many ontologies for knowledge representation, till now no one can be called universal, because it can replace the others. This is why knowledge graph theory was put forward and has been developed gradually. Chapter 3 focuses on the knowledge graph ontology and makes comparisons with a few other well-known ontologies. As knowledge is expressed by language the theory should be able to represent any language. The focus was therefore, in the beginning, on the representation of Chinese. Very specific aspects of Chinese turned out to be representable, see also Liu [Liu, 2002]. Also from our research we may conclude that the representation by means of knowledge graphs seems indeed independent of the language considered. ii) We propose how to express words by word graphs both semantically and syntactically in Chapter 4. In the knowledge graph project the focus was first on the



semantic word graphs only, so that we obtained an extension of the theory. We argue that the structural parsing developed in Chapter 5 of this thesis could be used both in English and in Chinese. For mapping a sentence on a sentence graph, both syntactic and semantic information is needed. As one of the goals is to develop translation systems, with the main steps: structural parsing, transformation of the sentence graph and uttering the sentence graph in the target language, an important result is that chunks can be used to develop computer programs for structural parsing. Given a sentence graph there are usually several ways how such a graph can be brought under words, i.e. can be uttered. Languages differ in the way the words occurring in the sentence are ordered. Chapter 6 studies the problem of determining rules for uttering a sentence graph both in English and in Chinese. The main conclusion from these three chapters is that indeed a translation system can be developed from knowledge graph theory. iii) Applications of knowledge graph theory are a challenge, especially in NLP. Based on the theory developed in this thesis, Chapter 7 develops a method for carrying out Information Extraction (IE). This was a third major item, that came along in a later stage of our research. Again the importance of considering chunks of sentences, and sentence graphs, should be mentioned. Here too the idea of structural parsing turned out to be fruitful.

ACKNOWLEDGEMENTS...................................................................... V PREFACE ............................................................................................. VII CONTENTS ............................................................................................XI CHAPTER 1 INTRODUCTION ........................................................... 1 1.1 NATURAL LANGUAGE PROCESSING .............................................. 2 1.2 A SEMANTIC MODEL OF NATURAL LANGUAGE PROCESSING ........ 2 1.3 OUTLINE OF THIS THESIS ............................................................. 4 CHAPTER 2 NATURAL LANGUAGE PARSING .............................. 7 2.1 INTRODUCTION ............................................................................ 7 2.2 DEVELOPMENTS IN PARSING ........................................................ 8 2.3 ASPECTS OF PARSING ................................................................. 11
2.3.1 Top-down parsing .............................................................................. 11 2.3.2 Bottom-up parsing .............................................................................13 2.3.3 Search techniques...............................................................................15

2.4 PARSING METHODS .................................................................... 15
2.4.1 Traditional methods ...........................................................................15 2.4.2 Parsing with knowledge graphs .........................................................17



CHAPTER 3 THEORY OF KNOWLEDGE GRAPHS...................... 19 3.1 FORMAL DESCRIPTION OF KNOWLEDGE GRAPHS ....................... 19
3.1.1 Concept ..............................................................................................20 3.1.2 Basic relations....................................................................................22 3.1.3 Relations in general............................................................................26

3.2 ONTOLOGICAL ASPECTS ............................................................. 27
3.2.1 Aristotle, Kant and Peirce ..................................................................28 3.2.2 Logic ..................................................................................................30

3.3.1 Fillmore’s case grammar....................................................................33 3.3.2 Expressing semantics with knowledge graphs...................................34 3.3.3 Structure is meaning...........................................................................36 3.3.4 Elimination of ambiguity in natural language ...................................37 3.3.5 A limited set of relation types ............................................................38

3.4 CONCLUSION.............................................................................. 39 CHAPTER 4 WORD GRAPHS: THE THIRD SET .......................... 41 4.1 INTRODUCTION .......................................................................... 41 4.2 CLASSIFICATION OF WORDS ....................................................... 42 4.3 LOGIC WORDS AND THEIR CLASSIFICATION ............................... 43
4.3.1 Classification criteria .........................................................................44 4.3.2 Classification of logic words of the first kind....................................50 4.3.3 Classification of logic words of the second kind ...............................53

4.4 WORD GRAPHS FOR LOGIC WORDS ............................................ 59
4.4.1 Proposition operators .........................................................................59 4.4.2 Modal logic operators ........................................................................60 4.4.3 Quantification ....................................................................................61 4.4.4 Logic words based on set comparison ...............................................62 4.4.5 Logic words referring to space and time............................................63 4.4.6 Logic words due to mental processes ................................................65 4.4.7 Words used in other logics .................................................................66 4.4.8 Words linking sentences.....................................................................67

4.5 CONCLUSION.............................................................................. 67 CHAPTER 5 STRUCTURAL PARSING............................................ 69 5.1 INTRODUCTION .......................................................................... 69

........7.2 Uttering rules from production rules..1 An introductory example......2 An example of representing patterns with knowledge graphs: KG-Structure.. 106 CHAPTER 6 UTTERANCE PATHS .......5 UTTERING GRAPHS CONTAINING FRAMES .................................................... 116 6.......................2..............3 Syntactic word graphs for word types .....6.............4 DESCRIPTION OF KG-EXTRACTION .......................77 5................ 87 5..............4.............3 UTTERING AN EXTENDED EXAMPLE SENTENCE GRAPH ................1 Definitions of syntactic and semantic word graphs .2 Word types for Chinese and English..................... 144 7..........2 SYNTACTIC AND SEMANTIC WORD GRAPHS .....134 6.............................................2......4 STRUCTURAL PARSING ....... 121 6.................................1 All........1 A traditional parsing approach ...........................4... 107 6........................................................................140 6..........................................2 Utterance paths and chunks ..........................7..................................2.............. 117 6.......................................125 6................................................. 110 6....4...............................................2..4 Examples of structural parsing....87 5...92 5...................138 6.......................1 Partial structural parsing ........4....3 Uttering paths for the extended example .... 133 6................................................................................72 5.....................151 .....Contents XIII 5.................1 Rules involving word types only................................7....88 5................................122 6........................................90 5....................4 UTTERING A SENTENCE GRAPH WITH REFERENCE WORDS .....................7... 143 7........4............................................... 113 6...........................1 INTRODUCTION ..................3 Chunk indicators ...6........................................................ 70 5.................................. each and every ....................................................................2 Rules involving phrases.............................. any.........................2 Uttering the word “dou1”......... 148 7..4......2....141 CHAPTER 7 INFORMATION EXTRACTION................150 7................................. 107 6.5 CONCLUSION..............................2 UTTERANCE PATHS AND GENERATIVE GRAMMAR . 143 7..............7 UTTERING AND GRAMMAR ................2 THE STATE OF THE ART OF IE ........................................................................7.. 150 7...................... 82 5..............................................70 5.............................................6 UTTERING QUANTIFICATION......137 6.3 OVERVIEW OF THE APPROACH .....3 GRAMMARS FOR CHINESE AND ENGLISH .................1 INTRODUCTION ..........

.............4........2.........4................................1 INTRODUCTION ..............2 ADWORDS ..................208 II............... 215 CURRICULUM VITAE......................................................2...................4................208 II.....2......2 The PAR-adwords ..........................................................3.......154 7...........3 WORD GRAPHS FOR OTHER PREPOSITIONS ........1 Adjectives ................7 A worked out example .................................. 199 II.............................6 Generating templates ........................................................................................................................................4.....2.............201 II.................................1...........154 7................204 II..........................................................1.....2........1 FPAR-classifiers .................. 192 I........4..4.. 183 APPENDIX I WORD GRAPHS: THE FIRST SET .....................2.. 199 II..........................1....................................................... 194 I...... 196 APPENDIX II WORD GRAPHS: THE SECOND SET ..............................179 BIBLIOGRAPHY ..............................................................2 WORD GRAPHS FOR SOME SIMPLE PREPOSITIONS .............4..............................................2......................... 213 SUMMARY .......3 Classifiers in the Chinese Language .................................................................................................1 INTRODUCTION ............................. 217 ..................................................................................... 191 I.......................8 Chunk graphs for the example ......1 The FPAR-adwords.....3................210 II..........................................................................210 II................. 191 I.............4 Automatic pattern acquisition ......................................154 7...163 7..9 Discussion .4 WORD GRAPHS FOR VERBS AND NOUNS ..... 211 INDEX .........................................XIV Contents 7...............................................2 Other classifiers .....................5 Inference and merging .............................152 7......................2..............................................1.................... 201 II....................................................153 7....................203 II...3 The CAU-adwords.......2 Adverbs......................................3 Named entity recognition..................2...............................................4 The ALI-adwords........206 II..................

all of these are based on knowledge. the research on knowledge representation propels the information age to change and develop from the elementary stage. to explore new methods for knowledge representation is still . Knowledge representation is a central topic in AI. With the development of the software for and the hardware implementation of computers. 1983. Artificial Intelligence (AI) has gradually got attention of more and more scholars. automatic programming. AI makes further applications to more and more fields. mainly with data processing. such as production rules. etc. logic. expert systems.Chapter 1 Introduction For more than half a century. see [Rich. its problems are not solved completely yet. Therefore. to the high level stage. Thus. or task describing. Although now there are many methods for knowledge representation. Graham. 1979]. or inferring and decision making. frame. information processing. automatic theorem proving. semantic network. a computer already can store much information and carry out fast information processing. natural language understanding. script. mainly with knowledge processing. machine study. Whether problem solving. robotics. It has important influence in fields like pattern recognition. and has become an interdisciplinary and frontal science subject increasingly. or expressing experience knowledge. During the last decades.

Knowledge graphs. expand the knowledge representation methods. Moreover. 1. as a kind of representation for NLP. and translate information. refine. as a new method. and is the thinking unit that refers to objective things and their peculiar properties. one determines a “general” set of properties that form the . information on INTERNET is growing unimaginably. knowledge graphs. but also to filter. Naturally. based on philosophy and psychology. to search for information automatically. There is a gap between formal languages (used by a computer) and natural languages. Considering various objects that are “special” cases. 1985].2 Chapter 1 one of the important subjects in AI.2 A Semantic Model of Natural Language Processing A concept is an important component of human thought. points out a new way for natural language describing and modeling. Fortunately. see [Rich. Harris. today. 1987. natural language (used by humans) is the most direct method and the symbol system most in use to express human ideas and pass on information. The approach to bridge the gap is often named natural language processing (NLP). This requires an intelligent information system. Communication between computers and humans is only possible when much research is aimed at bridging this gap. natural language understanding. or computational linguistics [Allen.1 Natural Language Processing Generally. 1983. The formation of a concept is a procedure that has the direction “from special to general”. The processing of these high level understandings must be and can only be based on semantics. It establishes a semantic explanation model for human perception and information processing. Graham. 1979]. and also makes a big step forward to the semantic understanding of “know it and know why”. on a high level of understanding. describing and modeling natural language is the base for the development of natural language understanding. and it determines the research process and the direction in the field of natural language understanding. 1.

This can be described as thinking is linking somethings. “boots” and “socks”. Certainly. the sentence graph. Therefore. . the course of human cognition is to establish “connection” for some concepts already in the brain. finally. So. The key of language information processing is to handle the meaning of a word. as mentioned in the above example. As the result of operating. lead to a larger semantic information graph. These smaller networks form a larger information network via connections. concept forms have a correspondence with the meaning of a word. these three are considered together. which is operational through the connection of word graphs. For instance. By connecting in the mind. When he meets a certain concept in language. In the other direction an object can only be described by the name of the concept if these general features are present in the object. the meaning of words and phrases has realized. he will think of every aspect that is related to this concept. etc. we will associate “wear the thing on foot” as the common feature of these three. we can associate the information related to its shape. through practice and study. where a concept is expressed by a word. Word graph formation (expressing the subjective meaning of a concept stored in a person’s brain) and connection operation (carrying out thought by a person’s brain). in order to form a network of related concepts. during he grows up. With the mind’s operation of forming concepts. the common feature of “boots” and “socks” is “be tube-shaped”. So. When we say “shoes”. Sentence graphs form the network of larger semantic information again through joining sentences. it is a kind of appropriate pattern of semantic understanding. a person establishes the concept of some objective thing step by step. which belongs to both the category of thought and the category of language. For instance. on different occasions or for different persons. Knowledge graph theory is based on the procedure of natural language processing that is assumed to be made by mankind.Introduction 3 components of the concept. The common point of “shoes” and “boots” is “be used for walking”. The essentials of the meaning of a word are determined by the perception of reality. color and taste. when we meet the word “apple”. after carrying out the joint operation continuously. It expresses a concept by a word graph. it generates the information network of a greater semantic piece. a difference in level and depth to understand the same concept is possible.

we propose a new kind of NLP technology. it is claimed that knowledge graph theory is original. Grammar rules are derived from the syntactic word graphs. The order of uttering words determines an utterance path. is discussed. chunk indicators are proposed to guide structural parsing. is introduced. As a result. which is called structural parsing. which can be used to map the sentence on a sentence graph. This is called structural parsing. In Chapter 6 the problem of uttering a sentence graph. The relationship with utterance paths is discussed.3 Outline of This Thesis In this thesis we study knowledge graph theory and its applications. On the one hand. we present a new class of words that are used in natural language. general and valid in representing knowledge.4 Chapter 1 1. Also it is posited that knowledge graphs are more general and more original than conceptual graphs. which is based on the theory of knowledge graphs. On the other hand. especially in natural language processing. due to the fact that the number of its relation types is very limited. In contrast with logic. the grammar rules are given for the Chinese version and the English version of syntactic word graphs respectively. In Chapter 4 we propose the concept of logic word and classify logic words into groups in terms of semantics. both semantic and syntactic word graphs are formed. By traditional parsing a parse tree can then be given for a sentence. Due to the distinctions between Chinese and English. and which are called logic words. The rules for utterance paths are investigated. bringing a sentence graph under words. In Chapter 2 the outline of parsing is given. We propose a . Furthermore we are concerned with its applications in natural language processing. this new method is further applied in extracting information from texts. A start is made with the building of the lexicon of logic words in terms of knowledge graphs. and some special problems on parsing are discussed. In Chapter 3 the basic theory of knowledge graphs is outlined from the point of view of ontology. In Chapter 7 we apply structural parsing to information extraction. Under consideration of the semantic and syntactic features of natural language. In Chapter 5 structural parsing. At the same time.

1993] are discussed.Introduction 5 multiple level structure based on knowledge graphs for describing template information. the direction of further research is focused upon. the relationships with the 10 functionalities mentioned by Hobbs [MUC-5. for Chinese. based on knowledge graphs. In the summary. . As a result. One of the goals is an automatic information extraction system.

6 Chapter 1 .

syntax and semantics.Chapter 2 Natural Language Parsing This chapter will aim at producing a general overview of some of the problems associated with parsing. The second section is essentially historical. . it is necessary to use a grammar that describes the structure of strings of words in a particular language.1 Introduction To understand something is to transform it from one representation into another. which is called parsing. This is done by a process. with the aim of describing the current state of the art and of outlining new research. A little bit of history of automatic parsing is mentioned. The first section covers some basic terminology and defines concepts of the theory. A sentence in a natural language can be analyzed from two points of view. To parse a sentence. covering early attempts to study aspects of parsing and recent developments in the field of syntax. The basic strategies of the parser involved are given. 2. and an evaluation of the different ways in which semantic information can be incorporated into a parser. The final section discusses the role of semantics within parsing.

on the other hand. a parser should be able to parse a sentence not only syntactically but also semantically. the sentence is perfectly grammatical. is concerned with the meaning of a sentence.” is syntactically incorrect. because we are interested in the meaning of sentences. On the other hand. Consider the two sentences a) “He saw the girl with a telescope. since the disambiguation comes from semantic information. Chomsky’s grammar theory influenced the field. it is likely to find a) and b) ambiguous in exactly the same way. with further elaboration in 1958 and . is concerned with the form of the sentence. However. Chomsky in 1956. or grammar.” Sentence a) is structurally ambiguous. Syntax allows us to identify word patterns without concerning about their meanings.8 Chapter 2 Syntax. etc. Hence. since the idea neither has a color. since the adjunct ‘with a telescope’ can be either a modifier of ‘the girl’ or an instrumental modifier of the verb ‘saw’. machine translation.2 Developments in Parsing In the earliest years of the research of natural language parsing.) require fast and robust parsing of large numbers of texts. If a parser has only syntactic information.g. The sentence “He saw any airplane. and we can easily analyze it into “determiner + adjective + noun + verb”. The sentence “The blue idea dreams. In general.” b) “He saw the girl with the red hair. 2. since English syntax states that the word “any” usually is used in “negative sentences”. information extraction. semantics is more important. many natural language processing applications (e. Semantics. Note that the syntax says nothing about the meaning of the sentence.” is meaningless if the words in it are given their usual interpretations. nor can dream.

including many that have never been encountered before. 1968. He developed a new theory. transformational grammar. it was Chomsky who was able to recognize the broader significance of these developments and to bring them into focus for the linguistic community. The argument transition network model of Woods [Woods. on rules that could generate the language. d}. Such information is easily available in the literature: the original 1970 article by Woods. first introduced the idea that languages could be interpreted as sets of strings. which represents a major theoretical reformulation for the field of linguistics. The grammatical formalism first introduced was known as the Chomsky hierarchy of grammar theory. who had already introduced the idea of “transformations”. Example Assume that the language is limited to a set {a. However. We show the procedure here: (1) S (2) S → a S a (3) S → A B (4) A → a (5) B → b (rule ii) (rule i) (rule iii) (rule iv) . see . c. usually abbreviated to ATN. 1982]. Chomsky’s approach was to shift from an emphasis on the language to an emphasis on the grammar. The rewriting rules (the grammar) can define an infinite set of possibilities. has enjoyed considerable success in computational linguistics. 1970].e. Then we can use the rules to generate a sentence “aaba”. Chomsky’s work had been preceded by the work of Harris [Harris. i.Natural Language Parsing 9 1959. Rules could be ( i )S → AB (ii) S → a S a (iii) A → a (iv) B → b. b.

It has provided a useful tool. by a desire to design ‘wholly semantic’ sentence-analysers.. 1975] on ATNs as a medium for linguistic descriptions. [Kaplan. This can also be illustrated with the introduction of formalisms like Lexical Functional Grammar (LFG). 1996] is evidential. 1975] and [Stevens & Rumelhart. in the early 1970s. [Waltz. in contrast to Transformational Grammar (TG) in its then Standard Theory (ST) form. 1968]). but this was followed.10 Chapter 2 [Bates. [Thorne et al. the Word Expert Parser (WEP) has been developed with particular attention paid to the wide variety of different meaning roles of words when analyzing fragments of natural language text. In the early 1980s. The ATN idea has had considerable influence in the field of natural language parsing. and text or discourse segmentation [Litman. 1993]. an important theoretical development has made the formal grammar theory become more spread. 1972] can be seen as a great progress in natural language parsing. part-of-speech tagging [Charniak et al. ATNs. [Pereira & Shabes.. on ATNs as front-end processor for database interrogation. For example. which explicates left-to-right processing. and is still an active one. 1997].g. 1975] on ATNs as model of human sentence processing. 1977]. summarized in [Woods. [Grimens. 1978] and [Woods et al. The more semantically-oriented work of Riesbeck. 1999]. 1965]. syntactic parsing [Ratnaparkhi. provided a set of actions offering a procedurally neat way to produce the deep structure of sentences. 1997]. and separates the parser from the grammar that is being applied. [Collins. there was an emphasis on purely syntactic parsers (e. [Kuno. 1989].. More specifically.. there has been a resurgence of statistical or empirical approaches to natural language processing since the late 1980s. 1992]. Generalized Phrase Structure Grammar (GPSG) and Functional Unification Grammar (FUG). [Charniak. . The success of such approaches in areas like speech recognition [Rabiner. [Manning & Carpenter. 1996]. 1972]. 1978] for a tutorial overview. On the other hand. These new linguistic theories stressed the role of the lexical information in automatic parsing. Wilks and Winograd [Winograd et al. Throughout the 1960s.

There are dozens of parsing techniques. two search techniques. The parse tree is a basic connection between a sentence and the grammar from which the sentence can be derived. the other is topdown parsing.1 Top-down parsing In top-down parsing. we start with the start symbol S and try to deduce the input sentence by constructing the parse tree. S → NP VP NP → the N NP → PN VP → V VP → V NP N → dog PN → he V → hit. means to reconstruct the parse tree that indicates how the given string can be produced from the given grammar. Suppose we have the following simple grammar for natural language. Also. 2.Natural Language Parsing 11 2. which describes how the grammar was used to produce the sentence. one is bottom-up parsing. a sentence. To reconstruct the parse tree corresponding to a sentence one needs a parsing technique. but only two basic types are reviewed in this section. according to a grammar. and suppose the sentence is “He hits the dog”. . depth-first search and breadth-first search. are mentioned in this section.3 Aspects of Parsing To parse a string.3.

The second one is fit for this sentence: S NP VP PN V NP . NP VP . We continue this process by applying the first rule for NP and the sentence is deduced by substituting the actual words: . The first rule would require “the noun” for some noun. the second rule would require a “pronoun”. this leads to the choice of applying the second rule and we obtain: S NP VP PN . We know that the production tree must start with the start symbol: S We only have one rule for S: S → NP VP: S .12 Chapter 2 First we try the top-down parsing method. Again two rules may be applied for VP: VP → V and VP → verb NP. We have two rules for NP: NP → the N and NP → PN.

When we find that the right-hand side of a rule can match with a segment. The first step is to recognize the word type for each word as follows: he hits the dog PN V the N . in which a sentence is deduced by using production rules from the left-hand side to the righthand side. we replace the segment with the left-hand side of the rule and repeat the process. until only the start symbol is left. Suppose we have the same grammar as above and suppose the sentence is also “He hits the dog”. . Note that we have to choose the proper rules to reach our goal. Now we try the bottom-up parsing method. Here the keyword is reduce.2 Bottom-up parsing In bottom-up parsing.Natural Language Parsing S 13 NP VP PN V NP the N he hits the dog . There is a search problem. 2. we start with the sentence as the input string and try to reduce it to the start symbol that is usually expressed by the symbol S. We reduce the input (sentence) to the substring (segment) that is the result of the last step by applying an inverse rule of grammar in postfix order. Top-down parsing tends to identify the production rules in prefix order.3.

So here we are forced to construct he hits the dog the N PN V NP .14 Chapter 2 Then we recognize the word “he” as derived by NP → PN. the word “the” and “dog” as derived by NP → the N. Hence we try he hits the dog the N . namely “V NP” that can been derived by VP → V NP. . PN V NP Again we find only one recognizable substring. NP VP And also our last reduction step leaves us no choice: he hits the dog the N PN V NP NP VP S .

which is used to represent the syntax structure of the sentence. 2. into a semantic model that represents the meaning of the sentence (independent of the context). and continue with that alternative until we reach a dead end. The distinction between breadth-first search and depth-first search is rather evident. which are depth-first search and breadth-first search in the production tree of partially generated solutions. Then we go back in the production tree to choose another alternative. Suppose there are several alternatives for further processing a partially solved problem. 2. a sentence is processed. unless we reach a dead end.4 Parsing Methods 2. Note that also in bottom-up parsing we may have more possibilities to choose from. In breadth-first search we keep all the alternatives for each partially solved problem. Generally. • In the semantic interpretation phase. using semantic knowledge. in which case we choose one alternative to repeat the same procedure. with arcs in opposite order. into a structural description (such as the syntactic parse tree).3.Natural Language Parsing 15 We have obtained the same parse tree. we can just choose one of both. in order to understand a sentence three phases of processing are distinguished. . In depth-first search we concentrate on one alternative. • In the syntactic phase. There are in general two methods for searching. Both parsing techniques can be used in structural parsing. Both of them are valid.4.1 Traditional methods Numerous parsing methods have been developed. the structural description is mapped. using syntactic and morphological knowledge.3 Search techniques Search techniques are used to guide the parsing through all its possibilities to find one or all parsings.

The main advantage of this approach is the simplicity of design of the program. which built a conceptual dependency structure [Shank.. • The integrated approach: the syntactic and semantic processing takes place simultaneously. Parsing methods can be roughly classified into three categories. The main distinguishing feature is the lack of a formal separation. The main disadvantage of this approach is that the semantic information is not used when a sentence is parsed syntactically.g. The main advantage of this approach is that parsing rules operate with both syntactic and semantic information. Perhaps the best known ‘syntax-first’ program was the question-answering system of Woods [Woods et al. [Kuno. • The non-syntactic approach: there is no syntax/semantics distinction. . 1974. and all the grammatical information is treated as being qualitatively the same. in the early 1970s. there was an emphasis on purely syntactic parsers (e. [Thorne et al. 1975a. and the second step to map this to a semantic representation. A simple example is parsing according to a semantic grammar. and a semantic structure is built directly from the sentence. One of the main issues in constructing a parser is whether to use an approach to separate syntactic and semantic processing of a sentence or to use an integrated approach.16 Chapter 2 • In the contextual interpretation phase. 1972. 1975] while scanning a sentence from left to right. 1972]. by a desire to have wholly semantic sentence-analysers. Throughout the 1960s. according to the way in which they handle syntax and semantics: • The syntax-first approach: the first step is to build a full syntactic structure for the sentence. but this was followed. throughout the parsing process. 1968]). One of the best known non-syntactic system was that of Riesbeck [Riesbeck. The main disadvantage of this approach is that it is impossible to construct parsing rules that can be applied to syntactic categories in general.. this representation is mapped onto a final representation of the sentence that is the real meaning of the sentence in the whole context. where syntactic categories are replaced by semantic categories in the grammar rules. and semantic information is used to limit the number of syntactic parses. 1975b]. 1965]. This can lead to a combinatorial explosion of the number of syntactic representations that are possible.

The syntactic processor converts a sentence into a syntactic sentence graph of the sentence. Thus a semantic sentence graph of the sentence is obtained. Moreover.e. .4. The lexicon is a list of syntactic word graphs and semantic word graphs. The semantic interpreter gives a semantic sentence graph that represents the meaning of the sentence. which is called structural parsing. partial meanings can be constructed as the sentence is processed from left to right) and utterance paths. The approach usually distinguishes two processors: a syntactic processor and a semantic interpreter.2 Parsing with knowledge graphs A new parsing method. In our structural parsing theory. we would like to set up a more flexible method that can both be used for the syntax-first approach and for the integrated approach. we pay more attention to semantic chunks (i. has been developed in the framework of knowledge graph theory. a syntactic sentence graph of the sentence that is processed can be constructed. Depending on this lexicon.Natural Language Parsing 17 2. based on syntactic word graphs and semantic word graphs. A very important knowledge source for structural parsing is the lexicon of word graphs.

18 Chapter 2 .

1 Formal Description of Knowledge Graphs Knowledge graphs. In principle. Ontological aspects of knowledge graphs are discussed by comparing with important other kinds of representations. It is expounded that knowledge graphs have advantages. to depict deeper semantic layers. while focussing more on the semantic than the syntactic aspects. belongs to the category of semantic networks. as a new method of knowledge representation. to use a minimum relation set and to imitate the cognition course of mankind etc. which is used to describe human language. the composition of a knowledge graph is including concept (tokens and types) and relationship (binary and multivariate relation). 3. which are stronger ability to express. Its appearance gave a new way to the research of computer understanding of human language. .Chapter 3 Theory of Knowledge Graphs Knowledge graph theory is a kind of new viewpoint.


Chapter 3

3.1.1 Concept
(1) Tokens Understanding human language lies in the perception of the real world (including general concept and individual instantiations). The thing that can arouse man’s perception in the real world is said to give rise to a token in the mind. In knowledge graph theory, we use the symbol □ to express this token. In fact, that people observe a thing is to signify there is such a thing in our real world. Therefore, in knowledge graph theory, everything will have a corresponding token. Definition 3.1 A token is a node in a knowledge graph, which is indicated by □. It expresses that we experience a thing in the real world or an existent concept in our inner world. (2) Types According to the viewpoint of subjectivism, different persons may describe experiences of the real world by different tokens. Some persons can experience some existent things in the world, that some other persons can not experience in the same way. So, in the inner world of a person, the token that can express existence may not exist in the inner world of another person. If a token can be shared by most of the persons, or can be shared by all persons, we have one objective picture. The most basic nature of perception is to divide different tokens into similarity classes. We observe that there is an identical type of the tokens that are in the same class, and that we can introduce types to express these tokens. For instance, if we see a dog or a tree, there is the token of dog or tree in the knowledge graph. So, we divide nodes into two categories that are types and tokens. In knowledge graph theory we may distinguish three kinds of marks. Mark □ shows a concept, and plays a role that is similar to the argument in logic. Type is the mark that is doing the labeling. It expresses the general concept that is determined by its property set. Therefore, type can be regarded as general information. Another kind of mark expresses the instantiation. That mark gives an example of the type, and it expresses the individual that is considered within the domain. In knowledge graph theory, we use the directed ALI relation between type and token

Theory of knowledge graphs


to express that the token has that certain type, and the directed EQU relation to express the instantiation. Definition 3.2 If a token is related to a mark by an ALI relation, which points to this token, this token is said to have the type expressed by the mark. The symmetric ALI relation expresses that a kind of thing and another kind of thing are similar. In graphic representation, the directed ALI link is used to point at a concept from a label, to give the concept a type with this label. However, A ALI B means that the concept B is similar to the concept A. Types of tokens are part of a type hierarchy. This hierarchy involves the well-known ISA relationship. If type 1 ISA type 2 then type 2 is more general than type 1. The most general type we consider is “something”, which may be seen as the name of the single token. “Type ISA something” holds for all types. An example should clarify the things so far. If the poodle Pluto is considered then we represent this by the knowledge graph




Here Pluto, □ and poodle are marks. We know that a poodle ISA dog, in knowledge graph theory the concept poodle, which is a graph, is considered to contain the concept dog, which is a graph too. The two graphs are related in the sense that the graph of “dog” is a subgraph of the graph of “poodle”. In the theory this is described by the FPAR relation, so dog ―FPAR→ poodle describes that poodle ISA dog. The FPAR relation will be discussed later. In most graphs that we consider the tokens are typed, i.e., there is an ALI-link towards that token. However, it may occur that a token without indicated type occurs. In “John hits”, the object is not mentioned. A transitive verb like “hit” is represented by hit

, in which only the central token is typed. The left token describes the subject and may


Chapter 3

e.g. be instantiated by “John”, but the right token may be left unspecified, knowing that John in an instantiation of a person, we would describe “John hits” by John




3.1.2 Basic relations
To describe the real world, we need to distinguish the relationships between tokens. In knowledge graph theory, the most important principle is to use a very limited set of relationships. These relation types are required to be basic. The more independent these types are semantically, the better. It should not be possible to deduce one relation type from others. The meaning of these relations can be described through considering the relationship with the real world. These basic relation types can be used for establishing more complex concepts and relations. How to choose these basic relations in knowledge graph theory? A basic relation is the relation of cause (CAU). In the initial stage of the knowledge graph project, this relation was the only relation type that people were interested in. In fact, in medical and social science, this one plays a very important role. Definition 3.3 The causal relation between two tokens is expressed by the arc labeled CAU. This relation expresses the relationship between a cause and an effect, or a thing influencing another thing. This relation type, not only in knowledge graph theory but also in other representation methods, is the relation that was distinguished most early, and is a basic relation which is used in a lot of inferences as occurring in a diagnosis system. The famous expert system MYCIN concerns IF THEN-rules only based on this one type. This relation points from the concept that produces the influence to the concept that has been affected, in the graphic representation.

Theory of knowledge graphs


There are also various situations in which the causal relation occurs. A thing or person can arouse an incident or course. An incident or course can also arouse another incident or course. An incident or course can arouse a state. Therefore, the complex structure that contains the causal relation is used to describe complex concepts, such as agency, purpose, reason, tool and result. For instance, in English the phrase “to hit with a stick” can be described as “to cause the stick to move resulting in a contact state”. When we discuss set theory, considering the relations between sets, we discover that four should be basic relations due to the following. Given “A” and “B” to express two sets respectively, we can distinguish the following relations: A = B , A ⊂ B ,
A ∩ B ≠ φ , as well as A ∩ B = φ . The four relations show respectively: A and B are

identical, A is a subset of B, A and B have common elements and A and B are completely disjoint. If we will regard A and B as the property sets of some special designated things, we must introduce relations to express these four relation types. Definition 3.4 A mark is a value, if it is connected to a token through a directed EQUrelation. An EQU-relation between two tokens expresses that two sets are equal. In graphic representation, this relation can express the concept naming through the arc from label to concept. This relation also can be used for thing assignment, for instance, red as value assigned to color. For a symmetrical relation, such as the equaling relation in set theory, we use the symmetric EQU-relation to join A and B. Special values are the perceptions by a person as individual, therefore, a mark is a very special value, and this special value can be expressed by an EQU-relation, this EQU-relation being directed from the mark that expresses the value to the token that indicates the perception. The reason for using the EQU-relation is that the special value is the assignment for the studied token. A very important relation is that a thing is a part of another thing. Definition 3.5 If there are two tokens that express two sets respectively, and one is a subset of another, then there is a SUB-relation between the two tokens. Note that there is a subtle difference between the SUB-relationship and the ISArelationship mentioned before. For a SUB b, there are two different interpretations.

that in which a and b are seen as property sets. place. When humans think about something they judge and ascribe certain attributes to things.8 The PAR-relation expresses that something is an attribute of something else. A DIS B expresses A∩B=φ. This led to including a new relation between an attribute and an entity.6 The ALI-relation is used between two tokens for which there exist common elements. like space.24 Chapter 3 • Concept a is a part of concept b.7 The DIS-relation is used to express that two tokens are in no relation to each other. tail SUB cat. The same holds for the EQU and the ALI relation. Definition 3. “the ball is red” indicates a relation like “red is the color of the ball”. As sets a and b are related by the SUB-relation. For instance. If the concept is expressed as a graph the elements of this graph will be said to be in an FPAR-relation with the graph as a whole. For instance. as are all other subgraphs of the graph representing b. Another relation that needs to be considered is the order relation. . Definition 3. “mammal SUB cat” expresses that cat is a kind of mammal and has more information than that involved in the general mammal. but as concepts. This relation expresses ordering with respect to a certain scale. length. therefore. we prefer to say that a is a property of b. For this relationship between a and b we used the FPAR-relation for its description. For example. this relation is from the attribute concept to the entity concept. Definition 3. In graphic representation. we have a contamination of terminology. Because of the symmetry. Note that we use a concept as a set of properties here. the DIS-relation is described by an edge instead of an arc. time. This relation expresses that a certain thing is attributed to or the external nature of another thing. In set theory. concept b contains at least all features of concept a. This expresses that the tail of a cat can be regarded as a part of the cat because the molecules of the tail form a subset of the molecules of the cat. represented by graphs. • Concept a is more general than concept b. In the second interpretation of a SUB→ b.

In information transmission. . this relation is a basic type of relation. respectively the verb and the object? That something is an object or a subject depends on its functional relationship with the verb. The meaning of the SKO-relation is based on the concept of informational dependency. But what is the type of relation between the subject and the verb. age. the ORD-arc usually points from the token with “low” value of the concept to the token with “high” value of the concept. which is called SKO (Skolem)-relation.Theory of knowledge graphs 25 temperature. but “what is said” is informationally dependent on the “saying” process. In natural language words like “depends on” are used. With this relation it is also possible to represent different tenses of language. functional or just mapping. as one of the basic relations that we must consider. but it also can be used to express “<” relation in mathematics. Definition 3. 1993]. etc. A change in the “saying” causes a change in “what is said”. we discover the mutual connection between information. We particularly refer to van den Berg [Berg. This relation is usually used for showing the order of time and space. Though in natural language there are many words to express mapping. we must consider a new relation type.9 The ORD-relation expresses that two things have a certain ordering with respect to each other. we choose to relate “man” with “hit” and “hit” with “dog” by a CAU-relation.10 A token in a knowledge graph has an incoming SKO-relation from another token if it is informationally dependent on that token. For informational dependency in mathematics we use the words function. In “man hits dog”. we still choose one relation type to express this relation. It expresses a relation as between “saying” and “what is said”. we use this relation. see Chapter 6. see also Section 6. we will be forced to add a new relation type. To express the dependency relation. In our concept world. The basic goal of the knowledge graph project is to use a limited number of relation types. by relating the time of an event with the time of speaking. weight. which corresponds to mappings. On the syntactic level we encounter a similar situation. This involves an aspect of choice. When comparing the order of two things. Only if the relation types are not enough to express something. When considering an ordering relation.6. Definition 3.

Hence. The FPAR relationship expresses that some subgraph of the graph is part of the whole graph that was formed. perception. see [Reidsma. feeling. 3. the relation between two concepts is also a graph. To express negation. Definition 3. Note that the contents of the frame may form the graph representation of a proposition.11 A frame is a labeled node. itself a frame. people were led to the three other frame relations. in the first stage of the knowledge graph project. namely some graph containing both concepts. In fact. Definition 3. To solve this problem. knowing. In the second stage of the knowledge graph project.14 NECPAR expresses the necessity of the contents of the frame. the FPAR-relation is the initial frame type. All nodes and arcs within the frame are connected to the frame node by the FPAR-relation. just like an arc can be interpreted as a binary relationship.26 Chapter 3 For that reason these syntactic relationships are also modeled by the SKO-relation. may also be modeled with this relation. Definition 3. the frame relation FPAR was introduced. At present in knowledge graph theory there are four frame relations in total. it is impossible to express everything in the world with only binary relations. A frame relation expresses that the labeled node is actually a frame around some complex graph. animal  FPAR→ cat. An example that should make .13 POSPAR expresses the possibility of the contents of the frame. Definition 3. which is used to express a complex concept or to express the word “and” in logic. and the possibility and the necessity in modal logic. Note that the graph can be interpreted as an n-ary relationship.12 NEGPAR expresses the negation of the contents of the frame. Apparently. is part of the “cat” graph.3 Relations in general Like the concept essentially is a graph. 2001]. three kinds of relation types are introduced. Of course. We already discussed the possibility to use a SUB-relationship here. The “animal” graph.1.

which is usually called an ontology. 3. represented by a node. It will be clear that different graphs containing a and b in principle determine different relationships. both the concept “John” and the concept “Mary” must occur in the graph that represents this sentence. of the token. up till now. In that graph two concepts occur that are the entities that are married to each other. homonymy is abundant in language.Theory of knowledge graphs 27 this clear is the relationship “married to”. There are many graphs possible for one concept name. For “married to” there are also many definitions possible. also called frame relationships.15 A relationship between two concepts a and b is a graph in which both a and b occur. However. . as is so obvious from the field of semantic networks. If “John is married to Mary”. Definition 3. It is also clear why there is a definite need for determining basic relationships. The eight binary types describe: ● ● ● ● ● ● ● ● Equality Subset relationship Similarity of sets. eight types of binary relationships and four types of n-ary relationships. The word graph ontology consists. This is clearly a concept in itself and therefore there must be a graph that is the meaning of this concept. alikeness Disparateness Causality Ordering Attribution Informational dependency : : : : : : : : EQU SUB ALI DIS CAU ORD PAR SKO. the number of types of relationships will grow indefinitely.2 Ontological Aspects We recall only the most essential parts for our discussion. If one does not pay attention to such a basic set of notions.

Once a subgraph of the mind graph has been framed and named another type of relationship comes in.28 Chapter 3 They are seen as means. Note that here WORDS come into play. The four n-ary frame-relationships are describing: ● ● ● ● Focusing on a situation Negation of a situation Possibility of a situation Necessity of a situation : : : : FPAR NEGPAR POSPAR NECPAR.2. Kant and Peirce Let us compare our ontology with two of the many ontologies proposed in history. that between the frame as a unit and its constituent parts. the relationships were considered to be on the sub-language level so to say. He distinguished: ● ● Quantity Quality ● ● Relation Location ● ● Time Position ● ● Substance Having ● ● Doing Being affected. in terms of awarenesses of somethings. 3. to structure the impressions from the outer world. on the level of processing of impressions by the brain. who distinguished twelve basic notions: . The focus there is on the way the world is built. These ten basic notions clearly focus on the physical aspects of the impressions. Any part of this graph can be framed and named. The situation is always to be seen as some subgraph of the mind graph. is called mind graph.1 Aristotle. a labeled directed graph in mathematical terms. available to the mind. The first one is of course that of Aristotle. as do the first eight notions of word graph ontology. using different types of neural networks. This structure. It will already be clear that word graphs for logic words will mainly be constructed using the second set of four n-ary relationships. The second ontology to consider is that of Kant.

The work of van den Berg can be seen as a direct continuation of this setup. was introduced by Peirce before 1900. 1993]. . Together with the “and” concept. including modal logic concepts like possibility and necessity. It has often been said that Peirce was guided by the ontology of Kant. which usually has an inner structure. Peirce’s definitions are not very easy to understand. From the point of view of knowledge graph theory the following stand is taken. we can distinguish: • The token itself. The two other frame relations give a way of describing all known systems of modal logic by means of knowledge graphs. something else. who presented the twelve basic notions in four triples. called existential graphs by him. see above. the conception of reaction with.Theory of knowledge graphs 29 QUANTITY Unity Plurality Totality QUALITY Reality Negation Limitation RELATION Inherence Causality Commonness MODALITY Possibility Existence Necessity Note that Kant clearly focuses on the logical aspects. 1885]. • • • Firstness: The conception of being or existing independent of anything else. which is simply two tokens framed together in knowledge graph theory. For any concept. Of course negation is included as well. whereby a first and a second are brought into relation. Describing logic by graphs. the definition of the concept.S. the negation gives a functionally complete set of logical operators for predicate logic. 1994]. Here some remarks are due concerning the work of C. secondness and thirdness of a concept. as was shown by van den Berg [Berg. token (or node) of a mind graph.Peirce [Peirce. when he introduced the notions of firstness. Secondness: The conception of being relative to. We quote from Sowa [Sowa. starting with the idea of simply indicating and (∧) and negation (¬) by two different types of frames. Thirdness: The conception of mediation.

background knowledge. In this view Kant’s triples do not correspond precisely to Peirce’s notions and we have the idea that the triple of knowledge graph theory: concept. 3. and r are predicates in symbolic logic. we know that any predicate formula can be changed into a . considered in relation to the concept.2. p p q q r In symbolic logic.2 Logic (1) The graphic view of first order logic Symbolic logic was started by mathematicians. suppose that p. q. possible for a word. The whole mind graph. the graphic symbol of each of them being given by itself. The variation in meaning. We list the graphic symbols and standard predicate formulas as follows: Graphic Symbol p ¬ ¬ ¬ q r Standard Logic Symbol (p∧q∧r) ¬(p∧q∧r) r ¬ (¬ ( p ∧ q ∧ r ) ) . Also the intension of the concept. also including what we call the background knowledge about the concept. is all that Peirce’s notions are about. is an intrinsic and essential part of knowledge graph theory. Peirce gave a graphic representation for symbolic logic. and has been applied in many fields. that we call the foreground knowledge about the concept. What is extra in our theory is the fact that the mind graph is not a fixed entity but depends on the particular mind (human) and for one human even on the particular circumstances in which the concept word is to be interpreted. foreground knowledge. For instance. its definition. is often not uniquely determined. inducing a subgraph of the mind graph. although it is one of the major goals in science to get at least the definitions straight.30 Chapter 3 • • The token together with its neighbors.

p∨q can be represented by the equivalent formula “¬(¬ p ∧ ¬ q)”.. Following we give the corresponding knowledge graphs. Besides. such as modal logic and tense logic. the disjunction. not only first-order logic but also other logic.. The connection words in proposition logic “and”. “if then”. Therefore. in knowledge graph theory there also exists the SKO-loop to express the universal quantifier. “not”. According to the above. the graphic symbol for standard logic of Peirce has established the logical foundation of both knowledge graph theory and conceptual graph theory. Extending the frame concept. etc. as well as the necessity and the possibility in modal logic etc. Since the universal quantifiers can be converted into the existential quantifiers in first-order logic. (2) Knowledge graphs and logic The FPAR-frame and the NEG-frame in knowledge graph theory correspond with the structure of Peirce’s graphs. the ORD-relation can express tense logic.6. For instance. have word graphs that are shown in Figure 3. Some complex predicate formulas are listed: Graphic Symbol ¬ Standard Logic Symbol p∨q∨r ¬ p ¬ q ¬ r ¬ p ¬ q ¬ r p→(q∨r) ¬ p q ¬ r s ( p ∧ q )→ ( r ∧ s ) . These kinds of graphic symbols can also be applied to conceptual graphs and knowledge graphs. can be expressed by these graphic symbols. conjunction as well as various predicate formulas derived from them can be expressed conveniently by graphic symbols. the graph of a logic formula is also called an existent picture. such as possibility and necessity. . “or”.1-3.Theory of knowledge graphs 31 conjunction form. including the POS-frame and the NECframe knowledge graphs are able to express modal logic.

The SKO-loop is used to express the universal quantifier. such as tense logic and modal logic. SKO Figure 3.3 Semantics in Natural Language Processing In this section. 3. so the theory has a very strong knowledge representation ability.7 “universal quantifier”. as well as reduces possible ambiguities. NEC Figure 3. Figure3.5 “possibility”. which expresses the meaning with structure. Figure 3. but can also express predicate logic and other logics. the problem of ambiguity in natural language processing is studied by knowledge graphs. At the same time. knowledge graphs can not only express propositional logic. Figure 3. . NEG BE NEG POS Figure 3. as Figure 3.3 “not”.2 “or”.32 Chapter 3 BE NEG NEG NEG NEG Figure 3.1 “and”.6 “necessity”. In a word. It is explained that the knowledge graph is a new method of natural language understanding. The existential quantifier is expressed by a distinct knowledge graph. depicts the semantic meaning from deep layers.4 “if…then”.7 shows. Somewhat special problems in Chinese natural language understanding are analyzed with knowledge graph theory. we study the knowledge graph method in natural language understanding.

manner. etc. present or future. EMPHATIC. “shi2 xian4” (implement) can be used both as a noun and as a verb. PROGRESSIVE) MODAL (CAN. MAY. aspect. MUST) ESSENCE (POSITIVE. modal. and does not discuss the modality part. INTERROGATIVE. The part-of-speech is not simply corresponding to the sentence composition. people often focus on the case grammar of Fillmore.3. For example. or must. IMPERATIVE) FORM (SIMPLE. This is determined by the properties of Chinese linguistics. form. how to describe the semantic structure of a sentence? In this respect. time. modal can be can. may. and such a connection is called a “case”. tense can be past. The modality part includes the concepts: tense. the important features of Chinese are the following: • • • There is no change in the shape of a word. • • • • • • • • TENSE (PAST.Theory of knowledge graphs 33 3. For example.1 Fillmore’s case grammar In natural language understanding. These concepts themselves can have different instantiations. NEGATIVE. INDETERMINATE) TIME (ADVERBIAL) MANNER (ADVERBIAL). The proposition part is formed by a verb and some noun phrase. . Fillmore focuses on the research of the proposition part. There are many special problems in Chinese natural language processing. PRESENT. Summarized. mood. The key of case grammar is that the deep layer structure of a simple sentence is composed of the proposition part and the modality part. FUTURE) ASPECT (PERFECT. The structure of a sentence and the structure of a phrase are consistent basically. essence. IMPERFECT) MOOD (DECLARATIVE. Every noun phrase connects with verbs by a certain kind of relation.

The knowledge graph can simulate this. and the concept “not married” in correspondence with another structure. as indicated by the drawn frame in Figure 3.2 Expressing semantics with knowledge graphs To establish a model for natural language understanding. developed by Yao [Yao. It puts in correspondence the “man” concept to a structure. connecting the two structures. a complete system for Chinese natural language understanding is the “lexicon semantic driven theory (LSD)”. Then. we can “expand” the concept “man” or “married” for further development to get a more complicated structure. from a psychological point of view. it is necessary to be able to express the meaning of a word or a sentence when the knowledge graph is used. Here. The word meaning is to be expressed through linking some concepts to other concepts. describes this semantic structure with a limited number of relations (8 + 4 kinds).8 has given a most simple kind of sketch that expresses this word by a knowledge graph. This is usually called the compositionality principle. it has to expand to more cases. to know the meaning of a sentence is to first know the meaning of each word. 3. the 49 kinds of relations can be expressed completely by the 8 kinds of binary relations and the 4 kinds of frame relations used in knowledge graphs.8 in its simplest form. Now. 1995].8. we obtain a new structure that expresses the meaning of “single man”. Therefore. This shows that knowledge graph theory is more basic for describing the semantic structure of a sentence. However. According to different requirements. This depends on the degree of . then gather all words into a sentence. The meaning of a sentence is a function of the meaning of each of its parts.3. A detailed comparison is considered by Liu [Liu. 2002]. which led to 49 kinds of semantic relations in total.34 Chapter 3 • There are some special sentence patterns. in order to know the meaning of the entire sentence. Note that Figure 3. If case grammar aims at describing the semantics of a Chinese sentence. in Chinese this is expressed by one word “dan1 shen1 han4”. In fact knowledge graph theory. We might connect the two concepts “man” and “not married”. Consider for instance the word “single man”. as indicated by the dotted frame in Figure 3. we first talk about the meaning of a word. He expanded case grammar further.

it is still a meaning of the corresponding word. Suppose that for the words that we use a lexicon of word graphs exists. different structures may come to his mind. This is a central feature of knowledge graph theory. The very simple graph for “married” just indicated that two tokens were involved. of that word.8 “single man”.Theory of knowledge graphs 35 complexity required. NEG married man ALI EQU Figure 3. The meaning of a word is not fixed. for which the lexicon has a word graph too. a single token. Each of the mental structures that the person can have in a correspondence with the word is a possible meaning. However. may lead to a collection of graphs that are very simple or that are rather complex. is replaced by the word graph in the lexicon that is given for it. Now. We should give a more thorough discussion of the idea of “expansion” of knowledge graphs. These numbers denote the four ways Chinese words can be pronounced. expansion means that some concept. It should be noted that this is also the case with . but we will not discuss this here. There are some problems here about the way such a more complex graph is to be included instead of the simple token. Making a lexicon of word graphs. Also from the psychological point of view this subjective aspect of meaning is relevant. This replacement leads to a more complex word graph for the concept in which it occurred. Also note that we have used numbers after the Chinese words. When a person ponders on some word. Of course the meaning of “married” is much more complex and involves various ways the two tokens are related. A more elaborate discussion of the concept indicated by the word gives more “meaning” to that word. interpretation. After the expansion the word graph for the concept considered has been enlarged. Then in these word graphs certain concepts may occur.

3 Structure is meaning Knowledge graph theory emphasizes that “the structure is the meaning”..36 Chapter 3 dictionaries and encyclopedia. Figure 3. 2002]. it literally says “borrow. borrow ALI CAU CAU CAU help ALI PAR EQU ORD CAU PAR ALI ALI time Figure 3. i. time 3. Without this analysis it is not quite easy to understand “borrow… to help self”. This word expresses something like “make use of”. Again. in the mind.e. for instance. To understand the meaning of “jie4 zhu4” an “expanded” graph is needed.9 shows. The meaning of a word is expressed by a word help self”. A certain word gets explanations of different size in these books too. in Chinese the word “jie4 zhu4”. this is even more so in ancient Chinese.. the meaning of a sentence is expressed by a sentence graph. has the meaning that Figure 3. The active “borrow” occurs before the action “help”.9 “jie4 zhu4”. knowledge graphs.3. The graph of a sentence is composed of the word graphs that express the meanings of words in this sentence by the operations of concept identification.. . In Chinese often word combinations force the listener to construct. As was found by Liu [Liu. rather complex pictures.9 can be explained by noting that the tokens for both the agent of “borrow” and the objective of “help” are the same.

the knowledge graph can restrict the redundancy of semantics considerably. In this way. as also in English.Theory of knowledge graphs 37 concept integration and relation integration. Figure 3. (presidium) tai2shang4 (rostrum) ALI PAR ALI zuo4zhe. These two different meanings can be expressed clearly by different sentence graphs.10 shows. the same meaning may be expressed by various sentences. another is “to be able to write well”. Consider for instance.10 Sentence graph of sentences (a) and (b). one meaning is “write well”. the meaning expressed by them is identical. Although these sentences look different. see [Bakker. Example (a) Tai2shang4 (rostrum) (b) Zhu3xi2tuan2 (presidium) ALI zuo4zhe (sit) zai4 (on) CAU zhu3xi2tuan2.3. (sit) presidium sit rostrum ALI EQU location Figure 3. the following two sentences (a) and (b).11 shows . What we meet here is that a sentence graph can be uttered in more ways than one. therefore their sentence graphs should have identical structure. This phenomenon will be discussed in greater detail in Chapter 6. so in knowledge graph theory they are corresponding to one and the same sentence graph. in Chinese the sentence “xie3 de hao3” has two meanings at least. the one that Figure 3. the deep layer meaning of these two sentences is identical. Though the structure of the surface layer of the sentence looks different. In Chinese. For instance. 3.4 Elimination of ambiguity in natural language In natural language there are many cases in which the meaning of a sentence is ambiguous. 1987].

If the set of relations is not restricted. 3. this relation will be added.3.38 Chapter 3 “write well”.5 A limited set of relation types In conceptual graph theory and in semantic network theory the number of relation types is not limited. This is the major difference between the two theories. In knowledge graph theory. man (AGNT) hit (OBJ) dog Figure 3. The conceptual graph and the knowledge graph are shown respectively in Figure 3.11 “xie3 de hao3” (1). We refer to [Willems. write ALI POS write PAR ALI ALI PAR ALI well well Figure 3. the types of relation are limited in number. The structure of these two graphs is different.14. We explain this problem in more detail with the following example.12 “xie3 de hao3” (2). and the ambiguity is not due to the sentence graphs. Each graph has a distinct meaning.13 The conceptual graph of “man hit dog”.12 shows the sentence “to be able to write well”.13 and 3. . 1993] for a detailed comparison of the two theories. Figure 3. there exists the problem of overlapping relations in semantics. Figure 3. Example Consider the sentence “man hits dog”. Whenever some type of relation is needed. The ambiguity comes in when the two different graphs are uttered in the same way. only the eight relations and four frames should be enough to express all semantics. A relation can be derived from another relation.

But in knowledge graphs. This is one of the notable advantages of the new knowledge representation method. the most important is its inference method. Whether an inference system is convenient. 3. For details on the inference problem. in knowledge graphs the relation set is limited. This chapter has not introduced the basic concept of the inference on knowledge graph. using only a CAU-relation. This fact claims that in conceptual graphs the structure easily leads to redundancy. then this token is an object. the knowledge graph is an appropriate semantic model of natural language understanding. Making a comparison between the two parts that are framed with dotted lines. However. 1997]. it is clear that Figure 3. but Figure 3. The meaning of a sentence. In fact.Theory of knowledge graphs 39 man ALI CAU CAU ALI dog ALI hit Figure 3.13 has used AGENT and OBJECT as two relations. both the agent and the object can been expressed. besides of the key part that concerns the logic and . then this token is an agent. so that the redundancy in structures is considerably eliminated. Therefore. namely the CAU-relation. The inference system of knowledge graphs has a very tight mathematical foundation. in conceptual graphs. see [Blok. and the OBJECT relation expresses the object (the object of the action in a sentence).4 Conclusion For a knowledge representation method. besides of its ability to express things. It should be the first selection for language information processing of researchers to express the meaning of natural language. as well as efficient.14 The knowledge graph of “man hit dog”. If a token is related to a CAU-arc that points out. directly affects its application level. the AGENT relation expresses the agent (the doer of the action in a sentence).14 has used only one relation. it has surmounted an obstacle for semantic networks in solving the complex problem of inference. If a token is related to a CAU-arc that points in.

40 Chapter 3 semantics of a sentence. speaker’s tone.1. It is part of the program to try to solve these problems with knowledge graphs.g. .3. mentioned as parts of the modality of a sentence in Section 3. also includes the extension of a sentence. such as e.

1996]. Word graphs were already built for prepositions and adwords (including adjectives.Chapter 4 Word Graphs: The Third Set This chapter deals with the third part of a series of word graphs in terms of knowledge graphs. which has its own word types. 1998]. and word structures. adverbs and Chinese quantity words) in two other papers [Hoede & Li. [Hoede & Liu. In this chapter. we propose the concept of logic word and classify logic words into groups in terms of semantics and the way they are used in describing processes.1 Introduction Natural language is kind of special symbol system that is used to express human ideas and pass on information. Each natural language has evolved into a kind of traditional symbol system. in which different components have some relationship and these related components play a role as a whole that has some meaning. This is why we study word graphs. A start is made with the building of the lexicon of logic words in terms of knowledge graphs. 4. A word is a basic unit in natural language processing. In science a particularly important part of the special symbol system is the set of words used in logic. It is therefore necessary to study the logic phenomena in natural language and the common rules in different languages in .

In particular in knowledge graph theory the meaning.2 we discuss aspects of classification of words. Its purpose lies in explaining the structure of language and the usage of a word. Therefore.2 Classification of Words It is necessary to classify the logic words before we study them. this is not the classification of grammar. graph structures. we try to reveal common properties of logic words that are independent of the particular language. In Section 4. In Section 4. This section is inserted because there are more ways to classify than the one used in traditional Chinese linguistics.3 the classifications of logic words are proposed. The classification of grammar is according to the grammatical function of a word. other logical structures and logical rules in sentences of a natural language text. of a sentence is a function of the meanings. In order to study the formation of words. Different purposes will result in different classifications. of its various parts. However. . which have a logical function in a sentence or even a paragraph. combining the sentence graphs we can obtain the meaning of the whole text. In order to study the structure of a set of single words it is usual to classify the words into monosyllabic. The first set and second set are given in Appendixes I and II. we classify words into simple words and compound words according to the number of morphemes.42 Chapter 4 order to reveal and explain pure logical forms. This chapter will study the set of words called logic words. By relating the analysis of linguistics with the analysis of logic. Finally. For example. to make this thesis self-contained. 4. In Section 4. How to classify a word depends on the specific purposes that classifiers have. in traditional Chinese linguistics. which is considered to be identical with the graph structure. to understand the meaning of a sentence one first needs to understand the meaning of each word that occurs in the sentence thus gaining the meaning of the whole sentence from these words and their order. disyllabic and polysyllabic words according to the number of syllables. two categories in the classification according to grammar are the notional word and the function word.4 we discuss word graphs for logic words. one of the syntactic aspects of the sentence.

which is according to some general purpose. “and”. In Chinese we have three main criteria. but not according to the narrow classification. The specific meaning attached to a word is affected by the philosophy. particularly in Chinese the meaning of a word strongly depends on the context. the social aspects. For the purpose of natural language processing. There are several criteria in a narrow word classification. “or”. The classification will be general. One is the word classification in the narrow sense. the main goal of knowledge graph theory is to construct a lexicon of word graphs. “not”. Words are considered: • • • in terms of the shape feature in terms of the syntax feature in terms of the meaning of the word. 4. Definition 4. Nevertheless.3 Logic Words and their Classification The words mentioned in traditional logic should be logic words. etc. Definition 4. “if… then…” and “if and only if”.1 A narrow word classification is a classification according to the grammar of traditional linguistics. another is the word classification in the general sense. Of these criteria the second one is considered to be the key one because Chinese has no distinct changes in the shape feature and the criterion of a word’s meaning may generate ambiguities. we investigate logic words and classify them.2 A general word classification is a classification according to the purpose of the classifiers. the ethical aspects. In line with the knowledge graph idea of meaning as a word graph. For example.. which are five words describing . which is according to traditional linguistics. involved in the discussion. the purpose being to make distinction according to the ontology of knowledge graph theory.Word graphs: the third set 43 In order to make a distinction between the purposes of traditional linguistic and general purposes we ought to divide the classification into two parts.

This class turns out not to be as easy to define and structure.1 Classification criteria Before discussing possible classification criteria we should mention the objects that we wanted to classify. “since”. the words. those words that are mentioned in traditional logic (including proposition logic.3. and to have a natural restriction for the set of words. For that reason. This class of words seems quite easy to recognize. . A corpus of 2000 English words with the property that they occurred most frequently in a set of 15 texts was established by Holland and Johansson [Holland & Johansson. deontic logic and fuzzy logic). such as “therefore”. “before”. are typical logic words. are related to logical aspects of utterances as well.44 Chapter 4 proposition logic. modal logic.. we considered these 2000 words and tried to classify them on a five-point scale: • • • • • definitely a logic word clearly related to logic. a) Subjective classification As a start of the general lexicon these 2000 words should be included. 1982]. “but”. The other logic words are words that are somehow related to logical aspects as well. many other words used in natural language. and then from a representation of the parts. So first the sentence is taken apart. Words like “possible”. “ought” and “permitted” which are used in modal logic and deontic logic are of course also logic words. Our main goal in this thesis is to develop a system of structural parsing. The pure logic words are then. etc. a representation of the sentence is constructed. But. 4. by means of which from a lexicon of word graphs the sentence graph of a given sentence can be constructed. We are therefore in need of more precise classification criteria. predicate logic. by definition. So the classification of logic words might include two parts. which are pure logic words and other logic words.” “while”. “necessary”. but not basic related to some form of logical reasoning having a vague logical flavor no relation to logic. tense logic.

The existence of two somethings. “probably”. C22. else. C35. equal. By functional completeness the other connectives from proposition logic follow from equivalences like p ∨ q ⇔ ¬ (¬ p∧¬ q) for the “or” connective “∨”. This prompts another way of classifying. impossible. put them in an FPAR-relationship with the frame. seen as two components of a frame. That frame can be named “and”. “possibility” and “maybe” as elements of this class. negation (seen as a quality by Kant). everything. “impossible”. “possible”. namely according to the occurrence of FPAR. “possibly”. C44. cannot. therefore. . POSPAR and NECPAR-frames in the word graphs of the word. nevertheless. if. every. C24. We would choose “might”. Similarly. C25. C13. C34. must. NEGPAR. existence.e. students in our case. hence. equivalent. ought. maybe. b) Classification by using Kant’s ontology Another way of classifying would be to use Kant’s ontology and decide whether a word belongs to one of his twelve categories. Note that these frames correspond to Kant’s categories existence. something in a NEGPAR-frame is put in a NEGPAR-relationship with that frame that now can be named “not”. C33. everyone. necessarily. doesn’t. then. true. thus. no. C23. necessary. difference.3 A logic word of the first kind is a word. that would fall in the category of “possibility”. right. the word graph of which contains one of the four types of frames in the knowledge graph ontology. possible.Word graphs: the third set 45 As a result we classified the words into classes C11. making a word graph for those words. use the POSPAR-frame in all cases. unless. C15. let us consider those words in the class C11 that we determined. possibility. expresses something of which the main feature is one of these twelve concepts. probably. As an example. C12. The two indices indicate the scale values mentioned by two classifiers. neither. might. “cannot”. truth. isn’t possibly. c) Classification by using knowledge graph theory ontology Looking at these words in b) from the knowledge graph point of view we discover that people. C14. Definition 4. i. nor. in order of the frequency of the words. wouldn’t. however. We give only the first class resulting from this subjective coding process. C11: and. possibility and necessity. C45 and C55.

The decision which word to use depends here on “cause” being a noun and “causes” a form of the verb “cause”. “B ke3 neng2 you2 A yin3 qi3”. S: A CAU B A POSPAR-frame around the whole of S would describe “(There is the) possibility (that) A causes B”. “cause A”. would lead to a description of “A (is a) possible cause (of) B”. literally “B possible have A cause” “A ke3 neng2 yin3 qi3 B”. “It is possible that S” or “Possibly we have S”. 3 or 4. A frame around A CAU . respectively. The reader should remark that the existential and universal quantifiers. literally “possible S” “S shi4 ke3 neng2 de”. say a situation S (German: Sachverhalt) in the form of a graph a POSPAR-frame may be considered around it. are not falling under Definition 4. In fact. A frame around B would describe. We may describe this by saying “S is a possibility”. is equivalent . “A possibly causes B”. literally “A cause B is possible of”. literally “S is possible” “you3 ke3 neng2 S”. In English “possible” is an CAU adjective and “possibly” an adverb. Subtle differences come forward due to the choice of the POSPAR-frame. In Chinese these three utterances are translated as • • • “ke3 neng2 S”. literally “A possible cause B”. Suppose S is given by the following graph.3. the three sentences are translated • • • “A yi3 qi3 B shi4 k3 neng2 de”. indicating the four forms of intonation.46 Chapter 4 When we consider something. “there exists” and “for all”. Peirce already pointed out that making the statement S on the paper of assertion (in whatever form). we have chosen to give the Chinese sentences and words in this thesis in spelling form. which by itself reads. As mentioned before. literally “have possible S”. In Chinese. in so-called total graph form in which the arc is also described by a vertex. followed by a number 1. pin1 yin1. 2.

etc. So “all” is not a logic word of the first kind according to Definition 4. the empty frame. “Exist” is a logic word of the first kind as its word graph is the “be”-frame. Other logic words. a subgraph of the mind graph is framed and in that way the definition of a concept C is given. that we like to describe as “linking of somethings”. expresses the FPAR-relationship between “animal” (part of the definition of “dog”) and the “dog”-frame. For the other logic words the classifying feature in their word graphs is chosen to be one of the other eight types of. The description is “C is a …”. We should note here that “all” is falling in the category “totality” of Kant’s ontology. So here IS in ISA is a logic word of the first kind too.Word graphs: the third set 47 to existential quantification (for closed formulae). but then pure logic is not the whole of thinking. filled with “something SUB world”. like in “A dog is a(n) animal”. which also is a category of Kant’s ontology. then the answer may be “Possibly because A”. The universal quantifier in knowledge graph theory is expressed by the SKO-loop on a token. If B is a possible cause for C and A a possible cause for B. The sentence graph SKO CAU ALI DOG ALI BARK is to be read as “all dogs bark” or “for all dogs (holds) dog bark(s)”. we see that we start with C. . so words with one of the eight binary relationships as dominant link in the word graph. An important example is “causality”. The famous “ISA”-relationship. In knowledge graph theory we would classify according to the occurrence of a CAU-arc. In the framing and naming process. That is why we put “There exists” between brackets in our example sentences. Here “is” is a logic word of the first kind too as it describes the FPARrelationship between C and its frame content. If we analyze this thinking process. The ALI-link. are not really expressing aspects of pure logic.3. for “alike” concepts. corresponds to concepts in the “commonness”-category of Kant’s ontology and determines a set of words like “like”. “as … as”. binary. basic in words like “cause” or “because”. relationships. that should be read as “for all”. In expert systems the question “why C?” can be answered if causations are known.

By linking these data C and A CAU CAU we obtain A C . This process B of linking is particularly interesting when the concepts are expanded. Given some statements. but none of which contains the goal graph G as a subgraph. of their frame. From the given statements expansion. Both in this general process of reasoning and in the case of expert systems. In trying to find the (answer) graph A one meets the difficulty of expanding the graph in the right direction. Definition 4. they are replaced by the content.4 Logic words of the second kind are words the word graph of which contains one of the eight types of binary relationships of the knowledge graph ontology as dominant link. i. that is embedded in the rest of the graph considered. The statements have to be well-formed closed formulae. but we are not interested in them here. “If a graph A is true then its subgraph G is true” is the rule in the first case.48 Chapter 4 and then note that we know that B CAU B . In natural language the thinking process is often described by non-well-formed statements that nevertheless correspond to certain subgraph of the mind graph that can be used in the description of the expansion process.e. “If A then B” and “If B then C” are ruleversions for natural language descriptions like “B because of A” and “C because of B”. and have found A as possible cause. combining of “true” graphs. We will see that this deviation from pure logic allows for dealing with some other linguistic aspects as well. The rule-version has the pure logical setting in which the pure logic words are used. with available further knowledge. The expansion process plays an important role in thinking. Expand all given statements as far as possible. It is for this reason of almost equivalence in description that it makes sense to speak about other logic words. a “rule based” version can be given. The basic process of reasoning is namely that whenever a graph is considered to be true. may lead to many answer-graphs A that are all true. and the problem to prove that some goal-statement G is true if the given statements are true. This replacement poses problems of its own. say in mathematics. then the way to find the proof may be the following. CAU . till a graph A is obtained that has the graph G of the goal statement as a subgraph. but in natural language we would use the word “so” (which by the way turned up in class C12): “A so G”. each of its subgraphs must be true (under specific conditions on the structure of these subgraphs).

truth is the fact that both statement and model are parts of the mind graph. unless one link is clearly dominant. These are not logic words. one for p and one for (a part of) the model.g. as in physics. The outcome of the comparison determines the truth-value of p.g. A second way of looking at “truth” is as an attribute of a framed part of the mind graph. There are two ways of looking at these words. “Truth” can then be seen as equality of certain frames. Finally. as in mathematics. and hence is a logic word of the first kind according to Definition 4. Word graphs with more than one type of binary relationship are to be excluded. as perceived. To determine which link is dominant we need some measure for dominance. e. used often in thinking about structuring the world. or. The model is also held to be a correct description of the state of affairs. The knowledge graph theory slogan “the structure is the meaning” is in line with the .3. or due to ideas in an “inner world”.Word graphs: the third set 49 We have given a restriction here by demanding that the relationship. Meant are those words that describe the linking process in its basic form. more close to what was said before. The statements describing axioms are not considered to be in need of comparison. one for p and one for the model. The truth values “true” and “false” are nothing but instantiations of the outcome of the comparison that may be replaced by. definitely words often used in logic. the model of a situation. the CAUrelationship. In them a SUB-link was clearly dominant. the 15 different word graphs for the Chinese word for “in” were given. In graph theory measures have been developed for the concept of dominance and they can be used to decide whether a word can be called a logic word of the second kind or not. where e. some discussion is due on the words “truth” and “true”. as is often done. That part may represent the content of contemplation. Such a perceived situation may be held to be true or false. First there is the comparison of a statement or proposition p with a model of the situation expressed by p. is a dominant link as otherwise all words would be logic words. In the first paper in the series on word graphs [Hoede & Li.e. For our knowledge graph view. axioms are simply considered to be true. what is happening is comparison of two mind graphs. which in twovalued logic is “true” or “false”. A preposition like “in” can therefore also be seen as a logic word of the second kind. 1996]. the numbers 1 and 0. i. whether this state of affairs is due to a presupposed “outer world”. for example.

with proper domain and interpretations.I The first set of logic words of the first kind.6. a standard example. A statement like “ x < 5 ” is considered to have no welldefined truth conditional semantics even when a model is given. Of all these only the universal quantification did not involve a frame. universal quantification can also be represented by frames. 4.50 Chapter 4 second way of looking at truth. the outcome of such a comparison is irrelevant. In so-called truth conditional semantics this comparison is stressed. We already discussed the operators from proposition logic and the quantification operators. The first set of logic words of the first kind corresponds to the four frames themselves and is given in Table 4. Also modal logic operators were discussed.2 Classification of logic words of the first kind Due to our chosen Definition 4. The truth of the statement is depending on the comparison.I. but was expressed by a SKO-loop and was therefore a logic word of the second kind. . However. The statement “it is raining outside”. does not need comparison with a model. As a major consequence of our stand even statements that are not well formed also have a well-defined semantics as far as the corresponding mind graph frames are well-defined. WORD GRAPH Frame Negframe Posframe Necframe WORD Inherence Negation Possibility Necessity PARAPHRASE be with be not be possible be necessary CHINESE WORD gu4 you3 xing4 fei1 ke3 neng2 xing4 bi4 ran2 xing4 LITERALLY primitive gender not possible gender necessary gender Table 4. as far as meaning attribution is concerned. because x is free. The structure of the part of the mind graph associated by the listener with the statement is all that matters. In our structural semantics. with the outer world.3. Any knowledge graph constructed by a mind as corresponding to the statement is the meaning of that statement in structural semantics.3 we would have to construct the word graph to decide whether a word is a logic word of the first kind. We will come back to this in Chapter 6.

Secondly. “he2”. which is used to describe the occurrence of an alternative.Word graphs: the third set 51 Remark first the use of the word xing4. fPf. is . and “necessary”. . possibility is circumscribed by “possible male/female”. Omitting p from the graph the word graph for the word “not” results. Omitting p again the word graphs for “possible”. where male/female only functions to express the two values for possibility. corresponding to a proposition p. The word “inherence” clearly expresses more than just “being”. The second set of logic words of the first kind has graphs containing one of the four frames next to other parts. for “negation”. Likewise the POS-frame and the NEC-frame allow expressing “possible p” respectively “necessary p”. any subgraph that is framed and named gives a concept with the subgraph as inherent property set. Consider a NEG-frame containing a proposition graph P. By functional completeness other connectives in proposition logic are expressible by the “and”-frame and the NEG-frame. A frame containing the frames fPf and fQf is the representation of p∧q. “gender”. result as POS NEC and . possible/impossible. then is can be described by “it is not so that p”. be contained in a frame. respectively “∧”. Let a graph P. Literally. “ke3 neng2 de”. In the knowledge graph formalism the graph P: P EQU ALI PROPOSITION would be given. simpler. Similarly. The word graph for “and”. the word “Fei1”. or simply by “being p” or even just “p”. or. In Chinese this is described by. We have chosen the words inherence and negation as these are two of Kant’s categories. Note that the word “negation” has a subjective undertone. “yu3”. “bu2 shi4 p”. which may be described by “it is so that p”. used in the context of logic. without specifying the contents of the two inner frames. literally must be translated as “not”. “bi4 ran2 de”. literally “not be p”. or simply by “negation p” or “not p”. or p AND q in natural language.

according to his norm system. We should take a stand with respect to other logic like tense logic. but also that obligation is just an attribute of something. describes the word “existence”. see Table 4. the word “dang1”. deontic logic. or fuzzy logic and the words used therein. The frame corresponding to “being” may contain the graph of something considered to exist in the world. attached to the act by the speaker. literally “existence gender”. say an attribute of an act. be a set of numbers.52 Chapter 4 Note that in Chinese the word “de” is used to express the fact we are dealing with an adjective. Further information on the “world” considered. Consider for example the word “obligation” in deontic logic. is used. We now have word graphs for the logic words of the first kind describing logical operators. It may be argued that it is a word like “possibility” or “necessity”.g. Note that for “if”. “cun2 zai4 xing4”. LOGICAL OPERATOR WORDS And Not Proposition logic Or If…then If and only if Predicate logic Modal logic Existence (of) Possible Necessary CHINESE WORDS yu3 fei1 huo4 ru2 guo3…ze2 dang1 qie3 jin3dang1 cun2 zai4 ke3 neng2 de bi4 ran2 de LITERALLY and not or if…then when and only when exist possible of necessary of Table 4. If we follow the first . so essentially the graph FPAR ALI WORLD .II The second set of logic words of the first kind. may be added. is used in “if… then”.II. the word “ru2 guo3”. literally meaning “when”. whereas in “if and only if”. which may e.

One example should suffice here. We follow the second argument and do not consider “obligation” to be a logic word of the first kind or even an logic word of the second kind. Words used in tense logic are words like “present”. meaning of combining two propositions that may have no further commonness at all. because the EQU-relationship is clearly present. labeled. where the label TYPE may be . will be discussed in the next section. “future”. They are expressible in terms of the ORD-relationship concerning time values. “Number(a)∧Number(b)” may be the formulation in predicate logic. but to be similar to a word like “beauty”. For the logic words of the second kind we choose the eight binary relationships. meaning “a is a number and b is a number”. A fuzzy word from natural language is for example “youth”. The word graph for “youth” will clearly be quite complex. “Fuzziness” basically involves unprecise information about values of something. the “and”-connective has the.Word graphs: the third set 53 argument. Concluding. analogous to the POSframe and the word would have to be considered a logic word of the first kind. Whether it should be seen as a logic word of the second kind. vertex. we might introduce an OBL (obligation)-frame. For those words that have word graphs very close to those for these pure logic words we take the stand that although frames are used.I and 4. The reason for this is that we can then consider subgraphs of the graph TYPE . The representation we choose is that of the total graph in which the arc is also represented by a.3. In that case we consider “obligation” to refer to a given set of rules.II.3 Classification of logic words of the second kind For the logic words of the first kind the four frames in the ontology of knowledge graph theory were chosen as defining entities. the logic words of the first kind are those mentioned in the Tables 4. The essential meaning of “both” is that the predicate holding for a and b is the same. so at most fall under logic words of the second kind in case the ORD-arc is really dominant in the word graph. “past”. the essential meaning is not expressed by the frame. different. “Both” is therefore not considered to be a logic word of the first kind. No specific frame or basic relationship seems present here. 4. The word “both” is used in expressions like “both a and b are numbers”.

1996]. and those of the form TYPE .III The words corresponding to the graph Some remarks are due here. Table 4. Table 4. “Bao1 han2”. is not used. Secondly. the Chinese words tend to extend the description considerably. we notice that for the ORD-link. whereas “being equal” is considered to be a synonym of “equality”. that the merological stand is that SUB. This explains why the two columns for Chinese words given are so different apart from the words for “equal” and “equality”. hence “containment”. First.54 Chapter 4 one of the eight types we distinguish. This hints at the fact that also the English word “containment” expresses more than just a SUB-link in a BE-frame.V. Note. in Chinese the alternative indicating word xing4. It is simply so that not every knowledge graph has a precisely describing word. We might even have left more places open in the table for “lack of words”. TYPE EQU SUB ALI DIS ORD CAU PAR SKO Causation WORD Equal In Alike Distinct WORD(+BE) CHINESE WORD Equality Containment Community Disparateness Ordering Causality Attribution Dependency ying4 she4 xiang1 deng3 li3 xiang1 xiang4 bu4 tong2 shun4 xu4 qi3 yin1 CHINESE WORD(+BE) xiang1 deng3 xing4 bao1 han2 xing4 gong4 xing4 cha1 yi4 xing4 pai2 xu4 yin1 guo3 xing4 te4 zheng1 yi3 lai4 xing4 TYPE .III gives the words corresponding to the whole graph of three vertices and the graph in which an encompassing be-frame is considered as well. “gender”.e. . a more complicated word graph is expressed.IV. For the type EQU the graph is considered to represent the word “equal”. as discussed in [Hoede & Li. i. Thirdly. within a BE-frame. Now let us consider the subgraphs of the form TYPE and the corresponding words. given in Table 4. given in Table 4. we could not determine an English or Chinese word for the PAR-link. literally means “around inside”..

respectively B with attribute A. . mixed up in the English language by the fact that the words “of” and “with” are used in all three cases. • Attribute A of B.V The words corresponding to the subgraphs of form TYPE .Word graphs: the third set 55 PAR and FPAR are the three merological relationships. respectively B with part A. respectively B with property A.IV The words corresponding to the subgraphs of form TYPE EQU SUB ALI DIS ORD CAU PAR SKO WORD Equal (to) With Similar(to) Distinct(from) To With CHINESE WORD deng3 yu2 ju4 you3… xiang1 si4 yu2 bu4 tong2 yu2 dao4 ju4 you3 … te4 zheng1 LITERALLY equal (to) have … similar to not equal to to have … attribute Table 4. WORD Of… CHINESE WORD … de yi1 bu4 fen4 LITERALLY … of one part Table 4. TYPE EQU SUB ALI DIS ORD CAU PAR SKO From By Of Dependent(on) cong2 you2 … yin3 qi3 … de te4 zheng1 yi3 lai4 from have … cause … of attribute depend on TYPE . like in the examples: • Part A of B. • Property A of B.

The authors consider the four words to be slightly different so that four different graphs should be presented. would be described by “part”. In Chinese many words are used to describe what. This is probably the reason why in Table 4. although the SKO-loop. Like for the different words for “in”. This was not so apparent from English. The EQU. if barking is part of the definition of “dog”. The barking of the dog. in English. we can make word graphs for all these words describing merological relationships. more concretely: • • • The tail of the dog. Also note that the CAU-relationship. It is quite remarkable that all these words used in Chinese to describe merological relationships. we say that y depends (is dependent) on x. x . by van den Berg and Willems. indicating something that is informationally dependent only on itself and hence can be anything. In a more elaborate description they can be uttered by “…de xing4zhi” respectively “ju4you3… de xing4zhi”. The beauty of the dog.56 Chapter 4 Or. Special attention should be paid to the SKO-loop that is considered. In Chinese three different words are used. “sou3you3”. and “any”. “each”. where only a few words are at our disposal. which fact supports our choice of only three merological relationships can indeed be modeled with the three types occurring in the knowledge graph ontology. which is due to the fact that they do not seem to be present in language. “ren4yi4”. “mei3ge4”. as semantically the FPAR-relationship is used to modal a propertyrelationship. We will come back to this in Chapter 6. clearly stands central. “attribute” or “property”. In Chinese the two graphs can be described by “de” respectively “you3”. to represent the words “all”. “every”. ALI and DIS-relationships are symmetric. FPAR and FPAR The graphs are worded “of ” and “with ” too.IV no words are given. Some people define a dog as “something that barks and sniffs”. “mei3ge4”. ALI y ALI SKO If y is a function of x. with chosen word “by” is somewhat out of line .

recognize the following six: Inherence. and that puts the ALI-link central. “You2 … yin3 qi3”. although probably with great similarity. The listener may yet understand the used word in the proper way by searching for other concepts that contain as part of their word graph the same part as envisaged by the speaker. Necessity. Negation. The single SUB-arc gives the word graph for . Possibility. literally means “by … cause”. Speaker is considered by listener to have used the word for a part of a concept to indicate the whole concept.4. The ALI-link might be called a first among equals as the process of concept formation is seen to result from discovering similarity between a set of perceived objects. Rather. for the speaker. Of Kant’s categories we can. 1996]. although the specific feature. Using the word for a part of its complete meaning. plurality and totality. “Reality” in our subjectivistic theory is an assumption about correspondence between our image in the mind and a presupposed outer world. We do not discuss the quantity categories of Kant: unity. there is the problem to determine the borderline of the logic words of the second kind. Like for the logic words of the first kind. with some difficulty. “The big bird landed on Schiphol airfield” uses “bird” for “something that flies”. and hence a logic word of the second kind by Definition 4. but what about “like”? “A and B are alike” is a rather clean statement. on which the alikeness meant by the speaker is based. “Existence” we consider as “being in the world”. so to be less basic than “being”. discussed in [Hoede & Li. The listener may see this part of speaker’s word graph for “bird” as the intended focus of his word choice and may see it as part of listener’s word graph for “plane”. Limitation and Causality. allows to use the word metaphorically. The point is that next to the ALI-link more structure is needed in order to give a proper word graph for the word “like”. but in “he works like a horse” the ALI-link stands central but in the sentence graph the words “works” and/or “horse” have to be expanded in order to localize the parts that are similar. we want to focus on commonness.Word graphs: the third set 57 with the other relationships that seem more basic as structuring relationships. is not yet part of the expansions. A word evokes different subgraphs in different mind graphs. one of the things a horse can do. The simple expansion of “horse” to “working horse”. namely working hard. that we take to be synonymous with alikeness. “Alike” is the word with word ALI graph . already enables the localization. This is similar to the situation for the word “in”.

Here one should remark that with the noun its. like “beauty”. In tense logic the word “past” clearly involves an ORD-arc to time of the speech act from the time of the act described by the verb. Apparently in our minds no sharp borderline is present to distinguish logic words from other words. The nouns indicating the categories in Kant’s ontology. One might even take the stand that all words that are not noun. Both “past” and “lower” may be considered to be logic words of the second kind. In “the lower floors of the building” the word “lower” involves an ordering too. but that implies that more complex word graphs than those existing merely of one of the basic types of arcs or parts of them are to be called word graphs of logic words. but for the fifteen different Chinese words for “in” we have fifteen different word graphs. however indicated. is often missing in a sentence. It will also be clear that the classification formed by subjective coding of two coders and leading to the classes C11 to C55 mixes up logic words of the first kind and logic words of the second kind as distinguished in this section. In “I say (at time t0) John worked (at time t1)” we have t1< t0 on the time scale. In fact in Chinese “can” and “must” are not considered to be verbs. The essence in “past” is as much the ordering as the fact that we are talking about time. “shi4”. as the ORD-arc stands central. it might be considered to be a concept that expresses something attributed by the mind. but now concerning spatial coordinates. verb or adword are forming the material used to link these basic words. no other verbs being present. would also belong to the set of basic words. it is increasingly becoming more difficult to classify the word as a logic word when the graph becomes more complex. or t0 ORD t1 in short notation. We say that the verb “work” has the past tense. instantiations are included. would have to be excluded as well and this poses a more serious problem. However.58 Chapter 4 “in” in English. The verb “be”. We would prefer to call “like” and the Chinese words for “in” logic words of the second kind. An objection to this stand is that there are verbs like “be”. The number “2” is an instantiation of the noun “number”.4. Likewise all plant names are instantiations of the noun “plant name”. But then. these are logic words of the second kind according to Definition 4. these verbs are not. as far as we already interpreted them in the tables. “can” and “must” that are clearly pure logic words. But then. A linking of concepts by one of the eight basic types of arcs is predominant in prepositions. called auxiliary verbs. That again brings forward the problem to determine a borderline. without reason. as for example “possibility” was considered a typical word of the first kind. named. . All plants in the extension of “plant”.

philosophy. the structure of the concept is the same.Word graphs: the third set 59 Due to our view on word formation a classification should be based on the concept. that is the same in the minds of a Chinese or an Englishman.1 Proposition operators AND: BE NOT: NOT . have very simple word graphs. In the following listing we distinguish types of structures of logic words. 4. 4. as knowledge graphs are specifically designed for representing the structure of concepts. Although representing a concept is affected by sociology. in first instance. Types of frames are indicated in the left upper corner. 1998] next to the many nouns and verbs that. then Chinese classifiers and adwords [Hoede & Liu. There may of course be language specific concepts. Whether we choose Chinese or English to express a concept. [Hoede & Li. we will try to give knowledge graphs for some potential logic words. just describing the word as type. We already considered as first set prepositions. it is independent of the difference in languages such as Chinese and English.4 Word Graphs for Logic Words For our discussion about the classification problem for logic words it is useful to give a survey of the actual word graphs. For the process of structural parsing we need to have word graphs for all words anyhow. Before taking a more firm stand on the choice of the classification principle. 1996].4. psychology and so on.

2 Modal logic operators POSSIBLE: POS NECESSARY: NEC The nouns “possibility” and “necessity” have word graphs.4.3 these are word graphs for logic words of the first kind. corresponding to “being possible” and “being necessary”. POSSIBILITY: BE POS NECESSITY: BE NEC .60 Chapter 4 OR: NOT NOT NOT IF … THEN … NOT BE NOT According to Definition 4. 4.

However.4.6. However. see van den Berg [Berg. seems more appropriate and we would have word graphs like PAR OBLIGATORY and EQU ALI OBLIGATION PAR UNBELIEVABLE EQU ALI BELIEVABILITY . . If “possibility” and “necessity” are seen as judgments. whether something is obligatory or believable is a subjective matter.Word graphs: the third set 61 Within the frames. 1993]. We follow van den Berg and Willems and represent universal quantification by an SKO-loop: ALL: SKO . but we would clearly get away from the formalism of logic. for attribution. Unevaluated tokens correspond to free variables. Note that there may be many graphs of the same structure only differing in the instantiation. a token has been vaulted. i. we will come back to universal quantification in Section 6. graphs. analogous graphs could be given for “obligation” and “believability”.3 Quantification Existential quantification is expressed by an explicit knowledge graph in which a variable is instantiated. If we would allow two other types of frames.e. All these graphs correspond to “there is an x such that…” by explicitly mentioning the “x”. strictly parallel to “normal” notation for logic. The adjectives are seen as instantiations of the nouns. analogous word graphs could be given. 4. Hence a PAR-arc. can be present and we have a formalism. according to “obligatory” and “believable”. that describe propositions.

equivalent.4. four based on set comparison. Note the central position of the EQU-link. like e. There are many words of which “sub” is a part. (1) The equ-link The basic word here is of course EQUAL: EQU . The only word for which we would like to give a word graph here is “true”. The SUB-arc is typically structuring entities. 1996]. the types are to be divided into three groups. as was discussed in [Hoede & Li. “Truth” is then again obtained by putting a BE-frame around the word graph for “true”. With a BE-frame we obtain the word graph for “equality”.4 Logic words based on set comparison As we distinguish eight types of binary relationships.62 Chapter 4 4. The graph of the statement compared with the graph of the model in which the statement is interpreted should be equal to make the statement true. Should . we would have eight classes of logic words of the second kind to consider. However. two based on the structure of space-time and two based on mind processes. This preposition was extensively discussed in [Hoede & Li. (2) The sub-link IN: SUB . There are quite a few words in which the word “equal” occurs. 1996]. Hence TRUE: GRAPH ALI PAR STATEMENT ALI EQU ALI PAR ALI MODEL GRAPH could be given as the word graph for “true”.g.

It occurs in the word graphs of word like “similar”. even if this means including verbs like “disrupt” or “discover”. all these words are considered to be so. whom we are inclined to follow. Words like “from”. or a word like “seem”. if we want to list all structuring words as logic words of the second kind. there are many words including “dis”. which is almost a synonym of “alike”.4. “above” etc. “to”. (1) The ord-link Here again we refer to the first paper on prepositions.5 Logic words referring to space and time For the reflection upon space and time we assume that two basic types of relationships suffice. . in which the ORD-link is the main link. When we want to gather all words showing structuring aspects we should include all “sub”-words in the set of logic words of the second kind and give predominance to the “part of” aspect. 4. (3) The ali-link The basic word is of course “alike” ALIKE: ALI . For the same reason. Having discovered the alikeness in the examples of a species the mind may form a prototype structure and frame and name this. “under”. We recall that this type of link may be considered the first among equals as it is considered to be the basis of the framing and naming process. is appropriate here. Like for “sub”. the specific purpose of the classifier is decisive. We mention the ALI-relationship as playing an important role in metaphors. “behind”. DISPARATE: DIS . disjoint sets.2. the ORD-link and the CAU-link.Word graphs: the third set 63 all these words be called logic words? Is “subset” mainly describing a set or is the word mainly expressing a “part of” relationship? In a general classification. according to Definition 4. “after”. (4) The dis-link The example of two sets with empty intersection. although the CAU-link may be a composite according to the philosopher Hume. “before”.

the word “causing” seems to come closest. We do not give such graphs but rather mention that these three graphs without the token for process. With a BE-frame around it we might have . “today” and “yesterday” are considered to be logic words of the second kind. “now” respectively “earlier”. “present tense”. during. and analogous triples like “tomorrow”. (2) The cau-link We do not have a good word for the graph CAU . past and future. The words used in temporal logic mainly refer to some kind of ordering. would be word graphs for “at a time after speech”. Like we said before.64 Chapter 4 all have word graphs in which the ORD-link stands central. respectively before the speech act. In a rather complicated way these graphs will be present when “future tense”. respectively “past tense” would have to be described. PAR ALI SPEECH PAR ALI SPEECH PAR ALI SPEECH ALI TIME ALI TIME ORD ALI TIME ALI TIME EQU ALI TIME PAR ALI PROCESS ORD ALI TIME PAR ALI PROCESS PAR ALI PROCESS are describing that the process occurs after. two values of time are involved. present. Interesting examples are the tenses of a verb. as “tense” refers to the form of the verb that describes the process. which is usually described by “later”. These words. respectively “at a time before speech”. “at the time of speech”. so in smaller form. the time of speaking and the time of the process described by the verb. The ORD-link typically occurs in word graphs for words that are used in comparisons.

4. which we use for describing the functioning of verbs. The SKO-link. (1) The par-link Next to the FPAR-link and the SUB-link. to other types of words like nouns and verbs. this is the third merological relationship. Although the ending “-ly” in adverbs refers to the attribution to verbs. however.Word graphs: the third set 65 the word graph for “causation”. If y = f (x) indicates that y is a function of x. but also has syntactic aspects. has such a syntactic aspect as well. The typical example for the SKO-arc is mapping or function. now in a different way. adwords. in that it determines subject and object (if present). 4. which is essentially described by: . see [Hoede & Liu. may be seen as word graph. Causal relationships are very important in building expert systems or decision support systems. It is structuring thought like the other types of link do. for “cause” and “effect”. is typically used for adjectives. The PAR-link. and hence to the PAR-link. in that it links certain types of words. Subgraphs of the total graph like CAU and CAU . 1998]. we are not inclined to say that the PAR-link stands central in adverbs and do not call these words logic words. This again sheds some doubt on the CAU-link being basic.6 Logic words due to mental processes The last two types are the PAR-link for attribution and the SKO-link for informational dependency. (2) The sko-link The SKO-loop was used for universal quantification. For this reason and because we chose the CAU-link in our ontology. First expert systems in medicine were rule-based systems where the rules were formulated in the “if A then B” form instead of “A causes B”. we say that y is informationally dependent on x. “something that is causing” and “something that is caused”. All three occur in the word graphs of prepositions like “of” and “with”. as known from mathematics. the words with word graphs in which the CAU-link stands central are considered to be logic words of the second kind as well. with BE-frame describing “attribution”.

The concept “function”. the important role played by the ORD-link makes us say that many of these words fall in the category of logic words of the second kind. for words with a fuzziness aspect like “youth”. the SKO-link stands central in the word graph for “function”.5 that words used in temporal logic would not be called logic words in our classification according to Definitions 4. A word that seems to stand central here is “extent”. We prefer not to do this. However. The value of this measure may . Like in the case of “functional” and “mapping”. There are other logics like deontic logic or fuzzy logic. In natural language words for which the SKO-link stands central in the word graph are rather rare. which is basically a measure for inclusion.3 and 4. Words like these are called logic words of the second kind as well. In that case words like obligatory would be logic words of the first kind. “old” or “somewhat” we do not have the structuring aspect that other logic words have. SKO .4.7 Words used in other logics We have seen in Section 4. Likewise.4. where according to our classification we would not speak of logic words as none of the basic types of n-ary or binary relationship plays an important role. for mapping of arbitrary objects to numbers respectively mapping of arbitrary objects to objects.66 Chapter 4 DEPENDENT ON: DEPENDENCY being described by: BE SKO . as ethical valuations are essentially subjective. although we have mentioned that for deontic logic an OBL-frame might be introduced.4. as mapping from numbers to numbers. 4. is used in mathematical language and has word graph: FUNCTION: ALI NUMBER SKO ALI NUMBER .

but the word is used in a broader sense here. type c). is the meaning. For this reason we would not speak of logic words in case of fuzziness. 4. i. This encompasses the word and all its associations. which is equivalent to (p∧¬ q). but what are the boundaries of the specific subgraph of the mind graph? The extreme stand was taken that the whole graph focused upon. Let p and q represent two sentences connected by the word “but”. The main problem that we encounter is to make a decision about the demand “stands central” with respect to a specific type of link.e. of the graph.3 and 4. “nevertheless” or “but”. These words are frequently used in a reasoning process. the word graph would be BUT: BE BE NOT and therefore we should classify such words as logic words of the first kind. Within the same sentence the word “although” links sub-sentences in the same way. classification by using knowledge graph ontology. that contains the word graph. 4. is the meaning. The problem is similar to the problem of delineating the meaning of a word.4. As this is equivalent to ¬ (¬ p∨ q).4 for classifying pure logic words and other logic words. Examples are “however”.8 Words linking sentences The last group of logic words is special in that they link sentences. In knowledge graph theory the structure. background knowledge. their word graphs would not have natural boundaries as well. or ¬ (p→q) in logical notation.Word graphs: the third set 67 be called the fuzziness of the inclusion. Then the meaning is considered to be “it is not so that p implies q”.5 Conclusion It will be clear for which type of classification we have chosen. In Section 4. If the same stand would be taken with respect to logic words. The quantitative aspect stands central. .4 we have tried to apply our Definitions 4. although two types of frames are occurring. With “inclusion” we may think of sets.

The words and linguistic elements remaining.68 Chapter 4 The taken stand is rather off line with the mathematical logical tradition in which it is the goal to give a meaning that is as concise as possible. one would say the essential meaning. . We have made choices in Section 4. exclamation signs. The smaller the graph.4. the fourth set. calling some words logic words and other words not. The question is: does the word graph contain the link like any other type of link or is its structure so that the link grasps the essentials of the meaning. Here some frame or some basic type of link may be embedded in a larger graph that may still be the smallest graph giving the essential meaning of the word. the easier it is to answer this question. In knowledge graph theory interpretation this means that the word graph given should be the smallest graph still considered to give the meaning of the word. to name a few. This unsatisfactory feature is accepted because for the process of structural parsing that we have in mind the distinction between logic words and other words becomes irrelevant. will have to contain graphs for linguistic elements like question marks. For our classification problem we face the problem that we would like to call a word a logic word or not.

There are two important points here. Grammar rules are derived from the syntactic word graphs. chunk indicators are proposed to guide structural parsing. By traditional parsing a parse tree can then be given for a sentence that can be used to map the sentence on a sentence graph. This is called structural parsing.1 Introduction A natural language processing system always contains a parser. which is based on the theory of knowledge graphs. In this chapter. the grammar rules are given for the Chinese version and the English version of syntactic word graphs respectively.Chapter 5 Structural Parsing Parsing is an essential part of natural language processing. which is a device that has a natural language sentence as input string and that produces a representation of the sentence when it is acceptable. both semantic and syntactic word graphs are formed. The relationship with utterance paths is discussed. Under consideration of the semantic and syntactic features of natural language. is introduced. Parsing is the process of structuring a representation of a natural language sentence usually in accordance with a given grammar. As a result. structural parsing. 5. Due to the distinctions between Chinese and English. one is that we require a representation .

which is our approach to structural parsing. are given in Section 5.70 Chapter 5 as an interlingua (inter-transmittal language) that is standing between the natural language sentence accepted and its access structure in a computer. Since based on knowledge graph theory. It is fortunate that word graphs were already discussed for prepositions. which is called syntactic information. Considering the semantic and the syntactic information of natural language. we develop semantic word graphs for the meanings of words and syntactic word graphs for the syntactic function of words. The former point is independent of the specific language. A sentence graph is built from word graphs. another aspect of word graphs is discussed. the latter is dependent on the specific language. Some examples are discussed. Section 5. based on them. 5. adwords and logic words.4 is the central section and contains the discussion of utterance paths and their relation with parsing by chunks.2. which just stand for meanings of the words contained in this sentence. namely the semantic and the syntactic representation of a word by word graphs. the other is that we require grammars with which the natural language acts in accordance. due to their advantageous properties in NLP. the other is the function of the words. This means that word graphs are at the base of parsing.3. Here. one is the meaning of the words that constitute this sentence. The structural parsing that is introduced aims at transferring (or giving a meaning to) this sentence. 5. This chapter will discuss the theory of structural parsing.2 semantic and syntactic word graphs are introduced and the grammars for English and Chinese.2 Syntactic and Semantic Word Graphs We are interested in word graphs in term of which a sentence will be analyzed. In Section 5. parsing is more special than traditional parsing methods and is called structural parsing. two pieces of information are necessary.1 Definitions of syntactic and semantic word graphs To analyze a sentence to obtain a sentence graph. We chose knowledge graphs as the interlingua. We give the definitions for semantic and syntactic word graphs. and then take Chinese and English sentences as examples to support our theory. .

which expresses the syntactic functions of a word.1 Syntactic word graphs for the pronoun “wo3” corresponding to “I”. which expresses the meaning of a word. [Hoede & Liu. 1996]. 1998] and [Hoede & Zhang.2. For example. 2001a] on word graphs concern semantic word graphs.1. where on the semantic level we would choose a CAU-arc. . We can express its functions with three different knowledge graphs as in Figure 5. and this pronoun has at least the following three usages: ● subject ● object ● attribute. to respectively from a verb V. We have chosen to represent the subject function respectively the object function by a SKO-arc. Definition 5. SKO ALI WO ALI V SKO ALI WO ALI V PAR ALI WO ALI N Figure 5.1 A semantic word graph is a word graph. V ALI SKO PAR SKO ALI WO ALI V N ALI Figure 5.2 A syntactic word graph is a word graph. in Chinese the word “wo3” means “I”. The three papers [Hoede & Li.2 Syntactic word graph for the pronoun “wo3”. which is called the syntactic word graph of the word “wo3”. We can also express the three syntactic functions with one knowledge graph like in Figure 5.Structural parsing 71 Definition 5. The word graphs in Figure 5. The possessive use of “wo3” is expressed by a PAR-arc. “me” and “my” in English.1 give the different syntactic functions of the word “wo3”.

3 Semantic word graph for “wo3”. 1993] introduced syntactic graphs. We obtain that the syntactic word graph of the word “ta1” is the same as for the word “wo3”. In English there is no problem. which will be discussed in the next section. 5.g. One of the reasons is that Chinese words show no change of shape feature of words. in English “he”. Syntactic word graphs are essentially semantic word graphs for word types.3 Word types are the types of words. have the same syntactic functions in a sentence. or “his”. In fact many other words. in this case pronouns. But if there are no word types in our lexicon. WO EQU ALI EQU ALI SPEAKER PERSON Figure 5. it is necessary that we first classify the types of Chinese words. Therefore. This depends on how many word types there are. Definition 5. the intended structural parsing is impossible.72 Chapter 5 The meaning of the word “wo3” is expressed by another word graph. which is called the semantic word graph of the word “wo3”. Consider another Chinese word “ta1”. We use the basic ontology of knowledge graphs for all word types.2 Word types for Chinese and English In Chinese the problem of word types is more complex than in English. the type of a word always changes according to the context. “him”. prepositions.2. so their functions can be expressed by the same syntactic word graph. but took prepositions and function descriptions like subject or object as labels of arcs. we do not need to reclassify the words. Willems [Willems. The question now is how many different syntactic word graphs there are. including e. There is no Chinese dictionary with word types till now. see Figure 5. It is a problem how to divide Chinese words into types.3. classified in terms of their syntactic functions. Let us consider the following two Chinese sentences: .

In sentence (1). in the sentence “ta1 zuo2tian1 lai2 le (He came yesterday. “you2yong3” is a verb. The types listed here are based on the syntactic functions of words. For example. 1995]. (Swimming is good for the health. a Chinese linguist. we do not need to mention the semantic aspect of a word and can just concentrate on syntax. on the one hand. should not occur on the syntactic level of word types. There is no change in word shape. in the sentence “zuo2tian1 shi4 qing2tian1 (yesterday was sunny)”. In our theory we represent semantic and syntactic features of natural language separately by forming semantic word graphs for semantic features of a word and syntactic word graphs for syntactic features of a word.)”. What about English? Just look at the same sentence. “le” indicates that the sentence is in past tense. Based on the 22 word types of Zhu. When we classify word types. our purpose is to build a grammar that is independent of semantic features and dependent only on syntax. We list them in Table 5. He reclassified each of the above 22 word types into many subtypes. Due to this. 71 word types were classified by Yao [Yao. because he has to pay more attention to the word types semantically in order to make natural language processing possible. literally “He swim past time”. This is why .I. and looks like an adverb syntactically.Structural parsing 73 (1) (2) Ta1 you2yong3 le. There is the word “swam” to express the past tense and another word “swimming” to be recognized as a noun. Typically semantic aspects. the same word “zuo2tian1 (yesterday)” looks like a noun syntactically. Although no dictionary mentions word types. like time or location. (He swam. is the same in these two sentences. the word “zuo2tian1 (yesterday)” here modifies a verb “lai2 (come)”.) The Chinese word “you2yong3”. “you2yong3” is a noun. there are many views about how to classify Chinese words. representing Chinese word types with pin1yin1 and also give the names of these types in English. literally “Swim has good for health”. but has different word types.) You2yong3 you3yi4yu2 jian4 kang1. In sentence (2). Chinese words are to be classified into 22 word types [Zhu. According to Zhu. it is considered by Zhu to belong to the type of “time”. On the other hand. 1984]. For this reason it is neither an adverb nor a noun.

I that refer to the syntactic view to classify word types. Due to this we can reduce the bigger set of word types into a smaller one. .74 Chapter 5 in principle we agree with the types of Table 5.I Classification of Chinese words according to Zhu. CHINESE ming2 ci1 dong4 ci1 xing2 rong2 ci1 dai4 ci1 shu4 ci1 liang4 ci1 shi2 jian1 ci1 zhu4 ci1 jie4 ci1 fu4 ci1 zhuang4 tai4 ci1 lian2 ci1 fang1 wei4 ci1 yu3 qi4 ci1 tan4 ci1 chu4 suo3 ci1 xiang4 sheng1ci1 qu1bie2 ci1 qian2 zhui4 hou4 zhui4 biao1 dian3 fu2 hao4 ci2 su4 ENGLISH noun verb adjective pronoun numeral classifier time auxiliary preposition adverb state conjunction orientation modal particle interjection location onomatopoeia comparative prefix suffix punctuation morpheme Table 5.

they belong to the word building problem. Fourthly. We leave these out as well. . This belongs to another problem area that we call sentence patterns.) Wo3men2 qu4 you2yong3 (Do we go for a swim?) Wo3men2 qu4 you2yong3 (Let’s go to swimming. such as “!”. because they are unique features in Chinese. playing a key role in reasoning. so we leave it to further research too. we do not think a “mark”. la! (We went to swim. “ba le”. although they play a very important role in expressing meaning. (3) (4) (5) (6) Wo3men2 qu4 you2yong3 le. Modal particle words are used for: ● ● ● ● statement question suggestion exclamation : : : : such as “de”.) Wo3men2 qu4 you2yong3 (It’s very nice that we are going to swim!). etc. Secondly. ma? ba. Let us give the four usages and example words. We cannot give corresponding English words. we know that prefix. We also try to express each sentence in English. according to the association with the modal particle word. “?”. “a” such as “ba”. as a first reduction. “le”.Structural parsing 75 From our view. Here we just consider simple sentences. “a”. “a” such as “ma”. modal particle words are very interesting ones. should be a type of word. However. Now we give four sentences to explain these distinctive usages. They have at least four usages. “…”. “ni”. They are very small words and are only used in Chinese. We will not discuss sentence patterns in this thesis. “ni”. conjunction is always used to describe a connection between sentence and sentence. especially in English. Thirdly. suffix as well as morpheme are very important linguistic concepts. “a” such as “la”. “le”.

the semantic word graphs can express this. We do not give a table. Therefore we classify Chinese words into 8 word types. Also note that e. Because we would not like our system to be too complicated to work with at the beginning.76 Chapter 5 Note that all the modal particle words happen to appear at the end of a sentence. given in Table 5. We therefore start with the simplest and basic situation. so that there is no problem in parsing. we just chose the main word types as our target to begin structural parsing. . CHINESE ming2 ci1 dong4 ci1 xing2 rong2 ci1 dai4 ci1 shu4 ci1 liang4 ci1 jie4 ci1 fu4 ci1 ENGLISH noun verb adjective pronoun numeral classifier preposition adverb SYMBOL N V adj PN num cl prep adv Table 5. In English we also chose 8 word types.g. The main idea is to first give the main word types and build a grammar based on these word types. but the “classifier” type is replaced by the “determiner” type. “ma” can both function as a question mark and as a tone word.II with the terminology in English as well as the symbols that are used in the word graphs. we want to make structural parsing theory as clear and concise as possible. As for distinguishing the meaning. Modal particle words can be cut off from the sentence.II Restricted set of Chinese word types. Finally. and later expand our system by refining the word types and the grammar rules in order to process more complex sentences.

as there are several possibilities for the way that arcs link nouns. and the deep structure of a sentence is to be expressed by its semantic sentence graph. for example. In syntactic analysis. a classifier classifies a noun. Our further analysis will be based on 8 syntactic word graphs for the 8 word types given above. 1993]. The Figures 5. If.2. .II. Note that in the syntactic word graph for a preposition the type of arc is indicated by T. Noun and verb have the most complicated syntactic word graphs.3 Syntactic word graphs for word types The surface structure of a sentence is to be expressed by its syntactic sentence graph. then in the graph for the noun N the used PAR-arc should also be included.Structural parsing 77 5. see the graph for “cl”.5 give the syntactic word graphs for the 8 word types given in Table 5. These graphs are constructed by expressing the various functions. they will be combined to construct the surface structure of a sentence.4 and 5. also see [Berg.

.78 Chapter 5 adj adj adj ALI PAR PAR adv ALI ALI PAR PAR ALI adv ALI N N N adj cl ALI PAR ALI PAR SKO PAR adv ALI ALI PAR T ALI ALI PAR SKO PAR ALI num V ALI ALI ALI V N cl cl ALI prep PAR ALI N PN prep prep ALI num num N ALI PN ALI ALI T T T T ALI N ALI PN PAR ALI N Figure 5.4 Syntactic word graphs for 5 word types of Chinese.

.Structural parsing 79 PN V PN prep ALI ALI T SKO V ALI PAR V SKO ALI V adv V ALI P AR ALI PAR ALI PAR PAR SKO SKO ALI N ALI V ALI ALI adv N ALI adv PN ALI SKO SKO PN N adv adj ALI ALI PAR PAR PAR PAR PAR ALI ALI adj ALI V ALI V ALI N Figure 5.5 Syntactic word graphs for 3 word types of Chinese.

.6 Syntactic word graphs for 5 word types of English.80 Chapter 5 adj adj ALI PAR PAR adv ALI ALI N N N adj ALI P AR ALI SKO SKO V ALI ALI PAR T ALI ALI P AR SKO P AR ALI ALI ALI num det V N det det ALI PN prep prep SKO ALI N prep ALI num N ALI PN PAR ALI N ALI ALI T T T T ALI PN ALI N num Figure 5.

7 Syntactic word graphs for 3 word types of English. .Structural parsing 81 PN V PN prep ALI ALI T SKO V ALI PAR SKO ALI V adv V ALI P AR ALI P AR V ALI P AR PAR SKO SKO ALI N ALI V ALI adv N ALI adv PN ALI SKO SKO ALI PN N ALI adv ALI PAR ALI PAR PAR ALI adj V ALI V Figure 5.

such as “hong2 hua1 (red flower)”. etc. “yi4 ke1 shu4 (a tree)”.) . a classifier.3 Grammars for Chinese and English We try to derive a grammar. which were given according to the syntactic functions of words. or “zai4 jiao4 shi4 li3 (in classroom)”. There were 8 word types in Chinese and 8 word types in English. First consider Chinese. V go.) Here the verb “xue2” combines with the noun “you2yong3” in the first sentence. The following sentences are examples: (7) Xue2 V (Learn (8) Xia4wu3 N (Afternoon you2yong3 N swim lai2 V come hen3 adv very zhun3 adv probably rong2yi4. So there may be grammar rules where on the right hand side we read V N or N V. an adjective. and the noun “xia4 wu3” combines with the verb “lai2” in the second sentence. Consider a verb.82 Chapter 5 5. rules in a grammar indicate in what order the words can be combined. “san1 tian1 shi2 jian1 (three days of time)”. It is obvious that a noun can be combined with an adjective.2. in which “ke1” is a classifier. Such a combination of words should be possible as far as the syntactic word graphs are concerned.) xing2. it also has many ways to combine with other words. adj easy. In terms of the word types. Take the following sentence as an example: (9) Ta1 PN (He gang1 adv just zou3. The above 8 word types have many ways to be combined in a sentence. from the syntactic word graphs. for English and Chinese respectively. Also a noun can be combined with a verb as a subject or an object. a numeral or a preposition. given in Section 5. such as an adverb. adj well.

VP→V and V→N as rules. are added. No verb is mentioned. we see a pronoun followed by a noun. Special rules are needed to generate NP. the PAR-arc from adj to N in the graph for adj is found in the graph for N as well. We decided to include the rule VP→N. We now describe the grammar rules that we can derive from the syntactic word graphs for Chinese. the prepositional phrase. guarantee that the corresponding word graphs. Together they can play the role of a predicate in a sentence. Considering all pairs of word types we obtain a set of rules that are such that. Not without reason the first rule of grammar is S→NP VP. In “He baker”. The general way to find the rules is to check whether two graphs can be coupled. For example.) In the sentence the adverb “cai2” combines with the noun “shi2 liu4 sui4”. (10) Ta1 PN (He cai2 adv only shi2liu4sui4.Structural parsing 83 The adv “gang1” combines with the verb “zou3”. X is chosen to be N as nouns and verbs are the dominant word types in language. Note that they are considerably different from the syntactic word graphs for English as we have to take into account very specific forms of sentence building. an inverse rule adj N→X. This tells us that the ordered pair adj N can occur in a sentence. From S →NP VP we can obtain this sentence by using NP→PN. or. see [Hoede & Liu. as we prefer because of our aim to develop structural parsing. N 16 years age. 1998]. like in the last example where no verb is mentioned. VP and also PP. can be coupled. To enable the generation of such a sentence we need a production rule like V→N. This example shows the possibility of a rule with adv V on the right hand side. when applied in a parsing process. that is to be expressed. to which various other parts of the sentence graph. a rather common feature in Chinese. Sometimes a noun also can be combined with an adverb. . For example. where NP can be seen basically as an N and VP as a V. considered non-well-formed in English. syntactic or semantic. There should therefore be a rule X→adj N.

and sometimes like a verb. VP AP V N (The man walking to the school gate is my friend. A prepositional phrase can be attached to a verb and should perhaps then better be called an adwordial phrase. 7.84 Chapter 5 When considering the type N and V in an inverse rule N V→Y1 or V N→Y2. 4. it is pleaded to consider adjectives. for Chinese.) zou3 V le. 6. This combination can also be used like a verb phrase: (12) Ta1 wang3 xue2xiao4 men2kou3 PN AP (He was walking to the school gate. 1998]. 2. we may have NP. VP or S for Y1 and Y2. In [Hoede & Liu. in standard order: 1. irrespective of the type of the word or partial sentence to which the phrase is attached. The whole part is considered to function as a noun in this sentence due to the use of the word “de (of)”. This point is explained in the following sentence shown. The underlined part that is the combination of an AP and a verb is used like a verb phrase here. An AP is sometimes used in a sentence like a noun. 3.) The underlined parts “wang3 xue2 xiao4 men2 kou3 (toward the school gate)” and “zou3 (walk)” are the combination of an AP and a V. S → NP VP NP → PN NP → N N N N N → adj N → cl N → num N → adv N . (11) Wang3 xue2xiao4 men2kou3 zou3 de na4ge ren2 shi4 wo3 peng2you3. adverbs and classifiers as special cases of adwords. In our grammar we will use the symbol AP. 5. We now list the rules.

are more often used to modify a noun in an English sentence. NP → V PN 13. VP → PN V 28. VP → V PN 27. V → adv V → V adv → AP V → V AP → VV 24. AP → prep N 15. V 23. N N → PN N → NN 10. that is a particular word type in Chinese. In English there is no word type of “classifier”. we replace the classifier type by the determiner type in the set of English word types. NP → V N 11. V 21. the following English grammar rules were found: . V 20. adj 30. Determiners. 9. adj → adj adj → adv adj → adj adv . So. AP → prep PN 16. NP → PN V 14. VP → V 17. in an analogous way.Structural parsing 85 8. VP → N V 26. NP → N V 12. VP → V N 25. VP → adj 18. adj 29. VP → N 19. V 22. Based on the 8 word types that we found. like “the” and “a”.

7. The main reason is that certain sentences. yuan3. 3. V 15. VP → V N 13. adj tall). 5. 9. V 16. 8.86 Chapter 5 1. Note that we had to use more rules for Chinese. AP → V N 11. that would not be considered correct English sentences. V 17. had to be parsable in Chinese. adj → VV → adv V → V adv → V AP → adv adj . Consider the following sentences in Chinese: (13) Tian1 N (Sky (14) Ta1 PN (He gao1. an adjective or an adverb can never function as a predicate in English. V 18. S → NP VP NP → PN NP → N N N N N N → NN → adj N → det N → num N → PN N AP → prep N 10. 6. adj high hen3 adv very lu4 N road gao1. As a final example. adj far) . 4. VP → V 12. VP → V PN 14. 2.

Structural parsing 87 The underlined parts are respectively an adjective “gao1 (high)” or “yuan3 (far)”. this representation can also be a knowledge graph.4. and a sentence graph expresses the meaning of a sentence with a structure. the semantic graph of a sentence. but that often only one syntactic graph is suitable semantically. which we can call structural parsing. 5. The goal of structural parsing.2. Note that usually many syntactic sentence graphs can be derived by the grammar.1 A traditional parsing approach The following procedure could be used: . Knowledge graphs for words can be both syntactic and semantic. A word graph expresses the meaning of a word with a structure.4 Structural Parsing Parsing in natural language processing is defined as transforming a sentence or text into a representation of that sentence or text.4 Structural parsing is the mapping of a sentence on a semantic sentence graph. Definition 5. is in principle obtained as follows: • • • A grammar is used to construct one or more parse trees for the sentences. However. A syntactic sentence graph is derived from syntactic word graphs using a parse tree. Parsing by representing a sentence with a knowledge graph is a new field. which includes the concept of a word graph and that of a sentence graph. In both cases they are used like a verb. unless there is essential ambiguity. a sentence graph is to be formed from the word graphs of the words that appear in this sentence. or the combination of an adverb “hen3 (very)” and an adjective “gao1 (tall)”. A semantic sentence graph is derived from the found syntactic sentence graph. as we have seen in Section 5. 5. A word graph is then a basic unit in natural language processing.

. that erupted in 1992. The graph is “brought under words”.3. Syntactic word graphs are combined to a syntactic sentence graph according to bottom-up parsing. The volcano. erupted in 1992. For this reason. The speaker chooses an order in which these words are uttered. The key to our parsing method lies in a discussion of utterance paths.88 Chapter 5 • • • • In a lexicon for each word a semantic and a syntactic word graph is given for each use of the word. that erupted in 1992. we would like to give our own parsing method. Corresponding with this order is an ordering of subgraphs of the sentence graph. 130 kilometers from Anchorage. Both traditional bottom-up parsing or top-down parsing can be used to analyze a sentence. 5. which leads to one or more parse trees. The set of grammar rules. In Alaska. Alaska. 130 kilometers from Anchorage. With such an ordering of subgraphs usually one or more paths can be indicated. that lies in Alaska. that erupted in 1992. depending on whether consecutive words have overlapping word graphs or not. discussed in Section 5. One of the most difficult problems for traditional parsing techniques is to get rid of ambiguities. that is adapted to our knowledge graph theory. We do not like to produce many syntactic sentence graphs for complexity reasons.4. We can produce many syntactic sentence graphs that make no sense. is used in traditional parsing.3.2 Utterance paths and chunks A sentence expresses a sentence graph. lies the volcano. Each syntactic sentence graph is transformed into a semantic sentence graph by combining corresponding semantic word graphs. lies the volcano. 130 kilometers from Anchorage. if we use the grammars listed in Section 5. 130 kilometers from Anchorage. the word graphs. We will make the concept of utterance path clear by an example sentence: • • • • The volcano. lies in Alaska.

see also the paper of Abney [Abney. The ordering “in Alaska” does not occur in the third sentence. In the . The four sentences can be described as an ordering of the expressed chunks.8 these paths can be read off as texts in the frames. In all four sentences for one sentence graph we recognize typical paths. Abney states that people tend to express a sentence in chunks of words. not linked in the sentence graph.8 Semantic sentence graph with display of “chunks”. In the simplified sentence graph in Figure 5. are expressed as what might be called chunks of the sentence. Note that “jumps” occur. consecutive chunks. The remarkable feature is that the six indicated frames. that” also. and “the volcano. and we see that chunks of the sentence graph are brought under words in some specific order. but could have been used therein. “130 kilometers from Anchorage” is one such path. “Erupted in 1992” is another path occurring in all four sentences. that occur as connected graphs in the non-simplified sentence graph. 1 to 6: • • • • 1→2→4→5→6→3 1→2→3→4→5→6 6→5→4→1→2→3 5→6→4→1→2→3 . 1991] on parsing by chunks.Structural parsing 89 The volcano 1 EQU CAU CAU that 2 erupted in 1992 3 CAU lies FPAR 4 in Alaska ORD SKO 5 130 km from Anchorage 6 Figure 5.

….” and “The triangle has a right angle. her. should be a chunk indicator.4. may. its sides are 3. he. IF-THEN-frame or POS-frame in Chapter 4. In the sentence graph this is expressed with a POS-frame. etc. such as: it. so that what remains for structural parsing is “I with a listing of all flights from Amsterdam to Beijing”. The words “she” and “her” are chunk indicators in the first sentence.2. in the second sentence a jump 3→4. OR-frame. and hint at a chunk. for which chunks of the sentence graph seem to be easily constructable. where the chunk is the whole sentence. They are used to avoid repetition of mentioning something. In Section 5. which are auxiliary verbs. In terms of knowledge graph theory. in the third a jump 4→1 and in the fourth there are two jumps: 6→4 and 4→1. We have discussed frame words like BE-frame. The BE-frame can be seen as a chunk indicator too. his. can. its circumference is 12cm”. the fact that chunks of the graph are expressed as chunks of the sentence leads us to want to read off chunks from the sentence. For example in the following sentence: “Can I have a listing of all flights from Amsterdam to Beijing?” the auxiliary verb “can” modifies the whole sentence. the. we had the possibility to cut the sentence into two sentences by . 5.3 Chunk indicators Our reasoning behind the choice of indicators is the following. that. 4 and 5 cm. NECframe. this. and modify the whole sentence. as expressing a sentence graph. such as: be. see Chapter 4. With the interpretation of a sentence. frame words. Consider the sentence: “Every woman thinks she raises children better than her mother. A problem for finding chunks of a sentence is that of finding begin point and end point of a chunk.4. where “the volcano” occurred as a chunk. the word “its” is a chunk indicator in the second sentence. must. The auxiliary verb “have” is essentially “be with”.90 Chapter 5 first sentence there is a jump 6→3. as a guide line we will try to find chunk indicators. she. As our goal is to construct the sentence graph from a sentence. NOT-frame. Now consider reference words.

We will not develop a general theory for answering these questions in this thesis. Summing up we list the chunk indicators as follows: • • • • • Indicator 0: Pairs of comma and/or period signs. as are pairs of period signs. in “from Amsterdam to Beijing” or in “in Alaska”. If chunks are recognized. like all sentences. e. such as in the following sentence. “She gets up earlier than John. If a preposition is met in a sentence. with respect to grammar.. Likewise we might replace “it” by “the triangle” and obtain three sentences. where the word “up” cannot be combined with “earlier”. in which also other points discussed in this thesis should become clear.” Prepositions are very useful in natural language and always link other words.Structural parsing 91 replacing “that” by “the volcano”. including auxiliary verbs. Indicator 2: Reference words. or a pair of comma and period sign.g. Indicator 3: “Jumps”. are clearly chunk indicators too. but will consider a few examples as an experiment. including prepositions. In structural parsing. we can give the graphs of these chunks by combining word graphs. they hint at a “jump”. These sentences. Of course comma pairs. indicating a whole sentence. Indicator 1: Frame words. Indicator 4: Link words. How to link chunk graphs into a sentence graph. are clearly chunks.4. we do not think complete parse trees are necessary. we link these chunk graphs into a sentence graph. in written language. How to make chunk graphs for the found sentence chunks. After that. Of course now there are three problems: • • • To what chunks of the sentence do the indicators lead. see Section 5. . Therefore they should belong to different chunks. If two consecutive words can not be combined.2. it hints at a chunk.

The indicator 2 is about reference words. Prepositions combine with the noun following. “The volcano” is therefore a “complete” chunk. there is no further chunking. we chunk the sentence by checking indicators. commas and period signs.4 Examples of structural parsing In this section there are two example sentences: one is in English.[The volcano]. Example 1 “The volcano. “the” and “that”.4. We get in this way the resulting chunks and sub-chunks: 1. There are two reference words.3. [. We will give a detailed analysis. “kilometers” and “from”.” Phase 1 The preparatory phase contains two parts. between “lies” and “in”. as there is no auxiliary verb in this sentence. [[130 kilometer][from Anchorage]]. These jumps cut sub-chunks into smaller sub-chunks. there are three jumps. are separate chunks: “that” is a sub-chunk. 130 kilometers from Anchorage. As for the indicator 3. “in”. 2. According to indicator 0.92 Chapter 5 5. [[that][lies][in Alaska]]. indicating the different phases. This takes into account indicator 4. Other reference words. one by one. which were discussed in Section 5. CHUNK1 CHUNKS 2. the other is in Chinese. there are no sub-chunks. As there are no further chunk indicators. First. like pronouns. as well as “erupted” and “in”. Next we cut chunks into sub-chunks according to the other indicators. We do not use indicator 1.4. we get four chunks directly. that lies in Alaska. and 4 CHUNKS 5 and 6 . A determiner combines with the noun following. There are three prepositions. 3. 3. “from” and “in”. erupted in 1992.

CH9 = CH1 CH10= CH2 CH3 CH4 CH11= CH5 CH6 CH12= CH7 CH8.9. semantic as well as syntactic word graphs should be listed in a lexicon. for all the words in this sentence.” We will combine syntactic chunk graphs into a bigger one when they can be linked syntactically.Structural parsing 93 4.10. from CH9 to CH12. ALI PN SKO ALI V Figure 5. Checking these chunks. like in Figure 5. CH1. The syntactic word graphs used are only represented with relevant arcs. We use a new number to indicate a chunk. . see Figure 5. Since the syntactic word graphs have been listed in Section 5. [[erupted][in 1992]]. CH5. we found that only CH2 and CH3 can be combined into one chunk. others allow no linking syntactically in this phase. Since there are no further linkings syntactically. which may be a combination of sub-chunks.9. we now will construct semantic chunk graphs by using the simple semantic word graphs given in Figure 5. Second. The following figure shows the combination of CH2 and CH3. CH2 CH3 CH4. here we indicate them with word type abbreviations. CH7 CH8. Note that we might have given expanded versions of these semantic word graphs.2.] CHUNKS 7 and 8. We construct syntactic chunk graphs chunk by chunk. CH6. The other arcs are indicated by dotted lines. So now the sentence is expressed like “.11 The syntactic graph of CH2+CH3.



96 Chapter 5 Phase 2 In order to make things clear.12 and describe this in more detail. The semantic chunk graph of “from Anchorage” can be obtained by identification of tokens as indicated by the EQU-link from the token in the preposition “from” to that in the noun “Anchorage”. we also have some subchunks which cannot be combined in this phase. We do not consider word changes like “lies” instead of “lie”. like that for the number 1992. To achieve this we need some background knowledge. Due to the fact that the used semantic word graphs are so simple. The semantic chunk graph of “erupt” is very simple. In the semantic word graph of “in”. the tokens are not specified. which is “the volcano”. we renumber all the chunks as follows: CH9 = CH1 CH10a = CH2 CH3 CH10b = CH4 CH11a = CH5 Ch11b = CH6 CH12a = CH7 CH12b = CH8. Note that we expressed “1992” with an ALI-arc to indicate the set nature of the time interval. the verb “lie” cannot be combined with “in Alaska” directly. With simple semantic word graphs. so the right hand token can be identified with another token. The semantic chunk graph of “130 kilometers” can be obtained from the simple semantic word graphs of “130” and “kilometers”. • • For CH9. This is introduced by expanding the simple semantic word graphs into more complex ones. • • • • . Now we give the formation of semantic chunk graphs in Figure 5. the semantic chunk graph can be given directly. so we have semantic chunk graphs of CH10a and CH10b separately.


Then CH10a and CH10b can be combined into CH10 as in Figure 5. because semantically “lie” and “in” are both related to areas.13.14 The semantic chunk graph of CH11. .14. We expand “from” with an ALI-arc linking to “location”. For “in” the tokens are given type “area”. CH11a and CH11b can be combined into CH11 as in Figure 5. We do this. For “lie” we add two FPAR-arcs from tokens of type “area”. and expand “130 kilometers” with a PAR-arc linking to “distance” that has two SKO-arcs both to “location”. DISTANCE ALI PAR ALI KILOMETERS SKO EQU ANCHORAGE ALI LOCATION ALI LOCATION EQU 130 ORD SKO PAR ALI NUMBER Figure 5. So.13 The semantic chunk graph of CH10.98 Chapter 5 Phase 3 First we expand “lie” and “in”. ALI FPAR SUB ALI AREA ALI AREA LIE EQU EQU THAT LIE CAU EQU ALI IN SUB ALI AREA EQU ALI AREA ALASKA AREA AREA FPAR ALI ALI LIE SUB Figure 5.

Now we get the result: S = CH9.15. Linking “in” to “erupt” with a PAR-arc is the only possibility to combine them. see Figure 5. CH11. unless there is a jump.15 The semantic chunk graph of CH12. have been omitted for reason of clarity. CH12a and CH12b should be combined into CH12 like in Figure 5. and we obtain the semantic sentence graph. Finally. Here the sentence has 4 chunks. CH12. There are two points to mention here: “erupt” is a verb. semantically “erupt” is related to a “location”. LOCATION ALI PAR ALI ERUPT CAU PAR SUB ALI SET ALI EQU ALI NUMBER 1992 ALI TIME INTERVAL Figure 5. it should have a CAU-arc coming in.Structural parsing 99 We expand “erupt” with a PAR-arc and a CAU-arc. CH10. we link the four chunk graphs from the left hand side to the right hand side. which have corresponding semantic chunk graphs. . where some arcs used in the analysis. like the ALI-arcs to “set” or “number”.16.

16 The semantic sentence graph for the English example sentence. . Now let us take a Chinese sentence as an experiment to test the 5 phases of structural parsing.100 Chapter 5 THE VOLCANO ALI 1 EQU EQU EQU CAU IN THAT EQU 2 CAU LIE IN CAU PAR ALI ERUPT SUB EQU 1992 3 ALASKA EQU 5 SUB SKO FROM ORD PAR ALI ALI DISTANCE ALI 4 AREA ALI LOCATION SKO ALI EQU LOCATION ANCHORAGE KILOMETERS PAR EQU 130 6 Figure 5.

The “wei4zhi4” of “shang4xue2” “fills” the unspecified token in “zai4”. like in the lower part of Figure 5. CH6 and CH7 are expressed by semantic word graphs. the sentence is chunked as follows. • • • In Phase 2. Then the “wrong” CAU-arc from XBDX to “shang4xue2” is cut. In Phase 3.17. because “ren2” does not occur in the expansion of XBDX. [zai4 XIAN de] 5. • In Phase 5.21. Note that expansion is needed for linking “zai4” with XIAN. according to chunk indicators. [de] 3.19. Both problems are solved by expanding the semantic word graphs. see Figure 5. syntactic chunk graphs are linked into bigger ones.Structural parsing 101 Example 2 Wo3 de xiao3 di4di zai4 XIAN de Xi1Bei3Da4Xue2 shang4xue2. 1. In Phase 1. in the chunk graphs “ren2” is not linked.19. There are two problems: “zai4” has an unspecified token. see the upper part of Figure 5. we obtain syntactic chunk graphs like in Figure 5. [Xi1Bei3Da4Xue2 shang4xue2]. [wo3] 2.18. . [xiao3 di4di] 4. For “di4di” and for “shang4xue2”.20. the semantic sentence graph is completed. In Phase 4. a lexicon of this sentence is given in Figure 5.) (I of small brother in XIAN of west north university “My small brother studies in Northwest University of XIAN. expansions are chosen as in Figure 5. using syntactic word graphs. The CAU-arc “looks for” “ren2” and therefore links with “di4di”.” • • In Phase 0. and “ren2” has been added for CH6. study.

102 Chapter 5 WORDS wo3 SEMANTIC EQU ALI EQU WORD ren2 WORD TYPES wo3 ALI PN shuo1hua4zhe3 ALI ALI PAR de 1. 2. xiao3 PAR N prep PN EQU PAR ALI xiao3 adj nian2ling4 di4di ALI SUB di4di N zai4 ALI wei4zhi4 XIAN EQU ALI wei4zhi4 prep XIAN ALI cheng2shi4 EQU XBDX N XBDX ALI da4xue2 CAU CAU ALI shang4xue2 Figure 5. N shang4xue2 V .17 Lexicon of Example 2.

18 Syntactic chunk graphs of Example 2. .Structural parsing 103 CHUNKS SYNTATIC CHUNK ALI GRAPHS CHUNK1 PN prep CHUNK2 PN ALI T ALI CHUNK3 ALI PAR ALI adj N ALI T T prep ALI N CHUNK4 T ALI prep ALI N ALI N CHUNK5 SKO ALI V Figure 5.

.19 Semantic chunk graphs of CH6 and CH7.104 Chapter 5 PN ALI T T ALI N CH6: CH1+CH2+CH3 ALI PAR prep adj ALI prep N ALI prep ALI N ALI V ALI CH7: CH4+CH5 ALI T T T SKO Bigger syntactic chunk graphs CH6 wo3 EQU ren2 ALI ren2 PAR ALI PAR EQU xiao3 EQU shuo1hua4zhe3 ALI ALI di4di ALI nian2ling4 de da4xue2 ALI EQU PAR EQU ALI cheng2shi4 XBDX CH7 zai4 SKO ALI ALI wei4zhi4 PAR XIAN wei4zhi4 Figure 5.

21 The semantic sentence graph for the Chinese example sentence. ren2 de wo3 EQU ALI PAR di4di ALI PAR xiao3 EQU ALI nian2ling4 de XIAN EQU EQU PAR PAR zai4 wei4shi4 ALI SUB CAU ALI shang4xue2 XBDX PAR ALI wei4zhi4 Figure 5.Structural parsing 105 ren2 ALI ALI di4di ALI di4di PAR ALI wei4zhi4 shang4xue2 ALI CAU CAU ren2 ALI shang4xue2 ALI CAU ALI CAU PAR wei4zhi4 Figure 5.20 Expansions for “di4di” and “shang4xue2”. .

Traditional parsing focuses on the syntactic aspects of language and then faces the problem of dealing with semantics. Structural parsing then is the mapping of a sentence on a graph. The idea turned out to be quite fruitful. . which has corresponding linking of semantic word graphs. The word graphs originally considered in the theory are called semantic word graphs. This led to the idea that chunks of a sentence have corresponding chunks of the graph. Semantic sentence graphs were obtained. for which a traditionally flavored theory was designed by Abney [Abney. 1991]. That this is possible. i. Sentences are uttered in “chunks”. we find that parts of the semantic sentence graph are brought under words in such a way that “chunks” of the graph are expressed. In knowledge graph theory the approach is from the side of semantics. With a chosen set of indicators a Chinese sentence and an English sentence were investigated. however.e. i.e. a traditional parse tree obtained with the derived grammars will allow corresponding linking of semantic word graphs to obtain a semantic sentence graph. which is done for both Chinese and English. is considered to be a point of strength of knowledge graph theory. This starts with word graphs for the words. that the preparatory step of obtaining a syntactic sentence graph need not be carried out according to the traditional approach. however. From the syntactic word graphs grammars can be derived. the approach makes very clear where the aspects of artificial intelligence are located. As syntactic word graphs express the way words function with respect to each other. which are to be combined into a sentence graph. By investigating so-called “utterance paths”. The meaning of a word or a structure is considered to be a graph. be it that background knowledge concerning the words had to be included by expanding the semantic word graphs. The new concept introduced here is that of a syntactic word graphs for a certain type of word. The most elementary aspects of mapping a sentence on a so-called parse tree were mentioned in Chapter 2. It turns out. the structure is the meaning. the sentence statement is interpreted within a model.5 Conclusion The standard way of parsing is to find a generation of a sentence by a grammar.106 Chapter 5 5. As we will see in Chapter 7. Truth conditional semantics involves a comparison with a model.

erupted in 1992”. Let us start with a very simple sentence like “man hit(s) dog”.e. the sentence graph is given in Figure 6. CAU CAU ALI ALI ALI man hit dog Figure 6.1. In this chapter the concept of utterance path will be investigated in more detail.Chapter 6 Utterance Paths 6. We investigate several sentences both in English and in Chinese. The sentences arising from these ways of uttering consist of words occurring in the sentence graph in a specific order. i. can be uttered. that lies in Alaska. Given a sentence graph there are usually several ways how such a graph can be brought under words. Languages differ in the way the words occurring in the sentence are ordered.1 Sentence graph for “man hit(s) dog”. Here we study the problem of determining rules for uttering the sentence graph. 130 kilometers from Anchorage. .1 Introduction We have mentioned utterance paths in Chapter 5 in the sentence “The volcano.

Second. But how can the noun phrase be recognized in the graph? There are various ways to find out whether a word is a noun. “man”. our considerations about utterance paths are language dependent. Uttering the graph in Figure 6. However. There are 3! = 6 ways to utter the graph. Then we may ask what the sentence looks like that utters the situation expressed by the graph. In English. which is. in other language other orderings. “man” is the subject (S). We want to stress that. other utterance paths occur. 1988]. be uttered as “man hit(s) dog”. the first utterance path is used. a lexicon of words may explicitly say that “man” is a noun. as is “dog”. starts by the rule S → NP VP. In Japanese the verb is usually put at the end as in the second and sixth way to utter the graph. Any grammar. NP for “noun phrase” and VP for “verb phrase”. as that ordering has developed for these two languages. who discussed test sentences like: —— can be a pain in the neck. The graph in Figure 6.108 Chapter 6 Suppose that this sentence graph is given. “hit” is the verb (V) and “dog” is the object (O) grammatically. in English. in English. we could use the method described by Radford [Radford. with production rules. First. In our simple example the NP is “man” and the VP is “hit(s) dog”. We might say: • • • • • • man man hit hit dog dog hit dog man dog hit man dog hit dog man man hit.1 must. English and Chinese are therefore often called SVO-languages. therefore. and in Chinese. If a word can be placed in the slot indicated by —— to give a sentence that makes . where S stands for “sentence”. This sentence starts with a noun. of course.1 should therefore start with “the noun in the noun phrase”. The six orderings are usually described by the syntactic function of the words.

because of the rule S → NP VP. Then. An intransitive verb would only have an incoming CAU-arc. The tokens from which and to which the CAU-arcs are coming respectively going.Utterance paths 109 sense. We find “man”. see the syntactic and semantic word graphs in Chapter 5. are considered to be grammatical. as only verbs are represented with the help of CAU-arcs. We therefore should point out that there is a third way to find out the word type involved here. For example * Peter today book reads. the word is a noun. As English is a SVO-language we know that now first “hit(s)” and finally “dog” has to be uttered. A token with an incoming and an outgoing CAU-arc can only be a transitive verb. He mentions that in Serbocroatian the four words {Peter. . that does not occur in English or Chinese. Also semantic concepts play an important role. and find the noun from which there is a CAU-arc towards that verb. book. It is. is not allowed in English. from the structure of the graph.e. What we meet here is a phenomenon. Such non-grammatical sentences are indicated with a star: *. In the graph we see that we follow the path from the token “man” to the token “dog” via the token “hit(s)”. Indeed both “man” and “dog” pass this test. however. today} may be put in any of the 4! = 24 possible orderings without changing the meaning and. but hit (s) does not. There only 4 or 5 of the 24 orderings are good. 1988]. For our utterance problem we now know how to proceed. what is particularly interesting. all these utterance “paths” are allowed. with our rule for uttering. does not lead to “dog hit(s) man”. not only the syntactic interrelation of words that plays a role. let us consider an example given in [Radford. The problem with this method is that the outcome comes from the human being. who has to decide whether the sentence makes sense. looking at the CAU-arcs. i. must be labeled by words that are nouns. To make our point clear. Note that the ordering of the CAU-arcs. The utterance path has been found for our simple example sentence graph. Find the verb. start by uttering “man”. read. This too is due to the way word graphs are used. This makes hit(s) a verb.

2 Utterance Paths and Generative Grammar A sentence is a linearly ordered set of words. as the person reading or hearing the four words reconstructs the sentence graph in a unique way. 6. book. again. read. In the parse tree phrases are usually easily recognized as subtrees. exchanging the positions of “man” and “dog” gave a semantically completely different sentence. 1988] calls the parse tree a phrase marker. We conclude that there is. in any order.2.e. We may generate one example sentence by the rules: .2 Sentence graph with the word set {Peter. i. namely the one given in Figure 6. hardly any utterance rule. The solution is that “Peter” cannot be read and “book” cannot read. today}. Let us consider the sentence “Peter read(s) (the) book today”.2. Also the word “today” can only be attached to the verb “read”. which puts these pronoun and noun in the position of subject and object. In traditional parsing a parse tree is generated following grammar rules. But here exchanging the position of “Peter” and “book” does not have any consequence. purely on semantic grounds. let us recall that there is just one sentence graph. for the meaning of the sentence. In the case of “man hit(s) dog”. Radford [Radford. ALI PAR today Peter EQU CAU CAU ALI book ALI read Figure 6. In order to discuss the relationship between utterance path and generative grammar. Just utter the four words. in this particular case.110 Chapter 6 We can give an explanation why the 24 utterance paths are equally well possible. The sentence graph is the same for all these 24 sentences and is given in Figure 6. This sentence graph is the meaning of any ordering of the four words that is admissible. can be generated by the grammar.

S NP PN Peter V read V N book VP AP adv today The resulting parse tree is: Figure 6. not a rule N → N det.Utterance paths 111 S NP VP V AP PN V N adv → NP VP → PN → V AP → VN → adv → Peter → Read → Book → today. the sentence “Peter read(s) (the) book today”. today}. for example. So we see that utterance paths must follow the rules of the generative grammar of the particular language considered. Suppose now that there is a language Χ in which there is only one way to make a sentence out of the four words {Peter. In English the determiner “the” must be uttered before the noun. Then we . and let this be.3 Parse tree for one particular uttering. if the sentence graph is brought under words by an utterance path so that an admissible ordering. the grammar has rules that justify the utterance path chosen. read. and there is a grammar rule N → det N. for the generative grammar. As a very simple example we compare the utterance “the book” with * “book the”. book. is obtained. It is clear that.

Note that these numbers depend on the grammars for these languages and are numbers of phrase markers/parse trees that correspond to sentences with identical meaning. Figure 6. again. The main conclusion for the problem of utterance paths for sentence graphs is that it is very difficult to give general rules for bringing a sentence graph under words. There is a corresponding difference in the numbers of possible utterance paths for the sentence graph. If only one permutation is admissible the number of rules is large. On the other hand we may say that the number of utterance rules in Serbocroatian. • • “Utter some word” “Let a word be followed by a next word”.112 Chapter 6 have one sentence graph. Language Χ only allows one ordering and must therefore have only a restricted set of production rules.2. that SVO orderings are more often used. • • “Start with this specific word”. but in Serbocroatian there are 24 uttering possibilities. 2. If all are admissible. To enable a generative grammar to generate all 24 possible orderings of the four words. 4 or 5. The ordering of the words is precisely prescribed. the number of rules for uttering a sentence consisting of these n words is very small. . It is not unlikely. In Serbocroatian there is high flexibility. n. and in Χ there is only one possibility. then. etc. In Χ the number of utterance rules is large. The ordering of the uttered words is not dependent on the structure of the sentence graph. for a European language. for this example sentence. is very small. Χ is a typical SVO-language. Yet it may turn out that some of the 24 possible utterings of the sentence graph are more often used than others. In the language Χ there is no flexibility at all. for some grammar. in English or Chinese only a handful. “Then let it be followed by that specific word”. it must have many different production rules. There are n! permutations of n words.

in the garden. When uttered this way.1 can also be uttered as “dog is hit by man”. We will therefore not focus on general rules. the verb has to be sandwiched by the two NPs also in this way of uttering. thus suggesting a CAU-arc going out from this token. Figure 6. with a very big stick. given a unique sentence graph. and the APs “in the garden” and “with a very big stick”.” . that is expressed by the verb that must follow the outgoing arc. This example is now extended to the following sentence: “The mean tall man hit(s) the poor small dog in the garden with a very big stick”. the “be” frame that can be considered to be present around any sentence graph. hit the poor small dog. with a very big stick.Utterance paths 113 6.” Or: “The mean tall man hit(s) the poor small dog.1 can only be uttered as “man hit(s) dog”. We want to see in which order these phrases can be uttered. Something that is doing the act. The two APs can occur on any of the eight places indicated by the symbol in : man hit dog . but just investigate whether in English or Chinese patterns can be found in the way admissible utterance paths reflect the structure of the graph. We start with the token for the subject noun in the graph and follow the path determined by the CAU-arcs. following the SVO-structure of sentences in English. can be brought under words as “by (something)”.” Or: “The mean tall man. the fact that it is the “man” who is the agent is expressed by the incoming CAU-arc of “hit”. hit(s) the poor small dog.3 Uttering an Extended Example Sentence Graph We have seen in the last section that the number of ways to produce a sentence. For example one might write “In the garden the mean tall man. and therefore omitted. In written language we will use commas and in spoken language pauses will be taken. This extends our possibilities to order the phrases together with the central verb. depends strongly on the language used. Next to the verb “hit” there are the NPs “the mean tall man” and “the poor small dog”. However. First we remark that the graph in Figure 6. using only these three words. The word graph for “by” can be taken to be: CAU . As the object “dog” is mentioned first. is now explicitly mentioned. in the garden. with a very big stick.

But then the subgraph corresponding to “very big stick” shows a clear utterance path. The APs are adverbial phrases for the verb. Like for the example considered in Section 6.2. The four phrases correspond to four subgraphs of the sentence graph. Let us call the subgraph induced by the two CAU-arcs the main structure. they attach to the verb. the main structure follows the SVO pattern. Note that “the” with word graph ALI PAR ALI EQU (3 times) and “a” with word graph element set are not mentioned in the graph as we wanted to avoid frames in the figure. but before the head noun. We recall that chunks were introduced. for the same reason. Its arcs are of the PAR-type. The rather arbitrary order in which these phrases can be uttered. in English. The determiner has to be uttered first. We already saw how the corresponding subgraphs are uttered. poor”. The AP “with a very big stick” shows an extra feature.4. Then we can say that. Most importantly. by pointing out that people tend to make pauses in speaking. the two adjectives can then be uttered in arbitrary order. that all four have tree structure. when pauses are used in spoken language. unless commas are allowed. is possible due to the fact that the syntactic function of the APs is fixed. apart from the positioning of the two NPs with respect to the verb.114 Chapter 6 There is high flexibility in uttering the four phrases. starting as the outermost token for “very” via that for “big” to that for “stick”. The phrases are extensions of the head nouns or the verb. as then one might write “the dog. The two NPs have corresponding subgraphs that are identical in structure. Before considering the phrases themselves let us give the complete sentence graph in Figure 6. These subgraphs are phrases that correspond to the chunks that we considered in Chapter 5. . Also the prepositions “with” with word graph PAR and “in” with the word graph SUB are not mentioned. we must remark that these subgraphs are uttered “in one piece”. The connecting preposition is uttered first and then the determiner. small. this can be understood on semantic grounds. One cannot utter one of its constituent words while uttering another subgraph.

Utterance paths 115 mean EQU (character) ALI PAR (height) ALI PAR EQU ALI man EQU tall CAU big very EQU ALI EQU PAR ALI PAR PAR size ALI PAR hit EQU SUB ALI (location) PAR ALI (location) ALI garden PAR EQU (measure) ALI stick (instrument) ALI (sticks) CAU small EQU (size) ALI PAR PAR ALI EQU dog (subjective) ALI EQU poor Figure 6. Words between parentheses do not occur.4 Sentence graph for the extended example sentence. .

“The volcano lies. “those”. whereas the article “a” does not refer to a particular dog. In uttering one could cope with the occurrence of two verbs by uttering two sentences. we have a graph like .”.” and “The (same) volcano erupted. “her”. lies in Alaska. 130 kilometers from Anchorage. etc. “that” and “him” all have a reference function. the repetition of “The volcano” can be avoided by referring to the once mentioned volcano by means of the word “that”. The most common word is the determiner “the”. Due to the symmetry of the figure we can choose either of the two verbs for the main utterance path and utter the other verb in the subsentence containing “that”. that hated him”. “the”. but in “the man hits a dog” we have reference to some determined man. etc.4 Uttering a Sentence Graph with Reference Words In language many times reference words are used. erupted in 1992. that lies in Alaska. that erupted in 1992. on a side line so to say. this reference has been omitted. 130 kilometers from Anchorage. “that”. “she”.116 Chapter 6 6. “He”. “who”. However. serve to avoid repetition of already known or mentioned constituent words in a discussion.5 A subject for two verbs. In a sentence like “He hit the dog. etc. “He” refers to some man.” This sentence could also be uttered as “The volcano. In “man hits dog”. In Chapter 5 we considered the example sentence “The volcano. In the sentence graph these references should be recognizable. “that” to that dog and “him” to “he”. we have multiple use of reference. “which” and pronouns like “he”. that refers to something in the context of the conversation. “him”. Words like “this”. The basic structure of the sentence graph is volcano ALI EQU CAU lie ALI ALI CAU ALI erupt volcano Figure 6. As the reference is described by the EQUlink. “us”.” In both cases the word “that” refers to the volcano and is used for economic reasons. “the” to a particular dog. “these”.

V and O. Once the graph has been replaced by more graphs. with repetition of words. we see that we need not have a tree structure. Note that we could have contracted two of the EQU-links. In the second case reference words can be used. EQU Also note that the subgraph man ALI is given as the word graph for “he”. or leaving the graph intact. reference links can be introduced so that the main SVO-structures are completely represented with tokens for S. which is considered to have the meaning “something like a man. the two vertical ones.6 Sentence graph for “He hit the dog. containing only one verb. that reads “He might . 1988].3. while transferring all information attached to one token also to the other. If a sentence graph contains more verbs. where we omit a discussion of the tenses. but then the reference words would not have been so clear. At the same time some aspects of representing tense will be discussed. We consider an example mentioned by Radford [Radford. Before uttering. In the first case more sentences have to be uttered. that hated him”. 6.5 Uttering Graphs Containing Frames In this section we discuss the uttering of sentences with sentence graphs that contain frames. the occurring phrases can be uttered according to rules as discussed in Section 6. Now the utterer can choose between cutting the graph by deleting EQUlinks. to whom is referred”. We essentially have a graph for two sentences. “He hit the dog” and “The dog hated him”. here for “he hit dog” and “dog hate him”.Utterance paths 117 man ALI EQU CAU hit ALI CAU EQU ALI EQU CAU CAU ALI hate dog EQU Figure 6.

The things we want to investigate can also be discussed for the sentence “John may have been walking”. te ] . The tense. denoted by tb to the end of the act. We place “writing a letter” by “walking”. In the sentence graph this leads to a graph of the following form: John EQU CAU ALI PAR walk ts John John EQU PAR EQU [ tb. The present tense is characterized by ts ∈ [ tb. described by the verb. We intend to gradually increase the complexity of the sentence in order to see how frames come into play. denoted by te. te ] . is taking place to the time ts at which the sentence is spoken. First we consider the very simple sentence John John walks. te ] denotes the time from the beginning of the act. In the sentence graph “be” is expressed by a frame. . can be expressed by relating the time at which the act. present. The graph now looks like be John EQU CAU ALI walk . where the time interval [ tb. The change that has taken place is the use of the auxiliary verb “be”.118 Chapter 6 have been writing a letter”. san4bu4. EQU CAU ALI The sentence graph is John walk . Now we consider the sentence is zai4 walking. Both English and Chinese sentences are considered. san4bu4.

now attached to the “be”-frame. to describe the imperfect tense. “John now walk” is the literal translation of the Chinese sentence. be John EQU CAU ALI walk PAR ts EQU ORD EQU tb . expressing time. san4bu4. In Chinese the use of the word “le” is quite typical. The sentence graph is identical with that of “John is walking”. see Section 5.2. The fourth sentence is John was walking (when (Zuo2tian1 wo3 you2yong3 shi2) I swam John zai4 yesterday). namely “now”. In Chinese “zai4” can be seen as an adverb. but has not yet been ended for sure. In the sentence graph of the first sentence only the time description changes. san4bu4 le. This is described by the graph construction used in the third sentence. As a third sentence we consider the imperfect past tense. As the walking act has taken place in the past. The “be” frame is put around the sentence graph of “John walk”. where only the essential part of the graph has been given. The sentence reads John John walked. It is in one group with words like “ma”.2. we describe this by ALI PAR walk tb ts EQU ORD EQU . but for the description of the tense aspects by “…shi2”.Utterance paths 119 Note that there is hardly any difference in meaning. The ORD-relation expresses that the act began before the time of speaking.

in Chapter 4. that “have” can be interpreted as “be with”. describing attributes. The fifth sentence is John John has yi3jing4 walked. This very concisely describes that the act has indeed been completed. with respect to the graph for “John was walking”. Again the only change. If the interpretation of “have” as “be with” is taken to hold universally. The knowledge graph representation is more closely following the Chinese language. The sixth sentence reads John has been walking (when I swam (Zuo2tian1 wo3 you2yong3 shi2) John yi3jing1 zai4 yesterday). the SUB-relation. there is the problem of identifying the part of the sentence graph corresponding to the word “with”. san4bu4 le. at this point. The Chinese way of expressing the perfect past tense is more consistent with the way other tenses are expressed. describing properties. the FPAR-relation. There are three merological relation types in the ontology of knowledge graphs. The problem is the use of the auxiliary verb “have”. in Chinese the word “yi3 jing1” expresses the completion of the walking act. The perfect past tense is posing some problem. after having completed the walking obtains this as a property.120 Chapter 6 The literal translation of the Chinese sentence is “John in the past now walk”. It looks that the use of “have” in English to describe the perfect past tense must be seen as a special linguistic development. One way of looking at the use of “have” is that “John”. We have seen. then in “John has walked”. However. san4bu4 le. This word acts as an adverb of time. in English. All three merological relationships are expressed in English by use of the words “with” or “of”. with respect to the sentence graph of “John walked” is that tb has been replaced by te. describing parts and the PAR-relation. is the . So the sentence graph looks like ALI PAR walk ts EQU ORD EQU te . The only change that has taken place.

there example by a. In the beginning of the second phase of the knowledge graph project words like all. namely a. So. Anything stated about like x can be described by a sentence graph containing x. Note the use of the two reference words “it” and “that”. The peculiar use of the verb “have” in English for describing the perfect past tense should also be considered against the fact that in rather related languages like Dutch or German “has been” is expressed by “is geweest” respectively “ist gewesen”. The choice for the POS-frame can be defended also by the fact that yet another way to express the sentence graph is “It is possible that John has been walking”. In Chinese the word “ke3 neng2” is seen as an adverb. In fact. which can be uttered as “something like x”. If x is instantiated. an open formula in logic. then the graph a EQU exists an x.2 we said that the existential quantifier is expressed by a distinct knowledge graph.6 Uttering Quantification Let us recall the way existential quantifier and universal quantifier were expressed in knowledge graph theory. In Section 3. If this statement is P(x) then the sentence graph describes P(x). The universal quantification poses some delicate and interesting problems. put around the whole sentence graph for the sentence “John has been walking”. (Zuo2tian1 wo3 you2yong3 shi2) John ke3neng2 yi3jing1 zai4 san4bu4 le. Combining the graph for P(x) and the instantiation we obtain a knowledge graph that can be uttered as ∃x P(x). one might say “John possibly has been walking”. . for ALI x can be uttered as ∃x. In a very simple case we may have a graph ALI x . a closed formula in logic. and consequently in Chinese the use of the word “yi3 jing1” for expressing the completion. This feature is completely avoided in Chinese. The auxiliary verb “may” is expressed by a POS-frame. in those languages in stead of “have” the auxiliary verb “be” is used.Utterance paths 121 replacement of tb by te. or as “there is an x for which P(x) holds” in natural language. Finally we come to the seventh sentence John may have been walking (when I swam yesterday). any. 6.

per.6. 1988]. The logical statement ∀x P(x) is . regarded separately : 1. every 2. whole amount 2. 1984]. various subtle differences for the words used in universal quantification came forward. in the Oxford Pocket Dictionary [Allen. Words used in Chinese in universal quantification are dou1 suo3you3 de ren4yi4 ge4ge4 mei3ge4 : adv. in principle for the four mentioned words there should be four different word graphs.. no matter how much or many or of what : every one of two or more persons or things. all possible. some. let us consider the Chinese translation possibilities. all persons concerned. Before discussing these entries in the dictionary. one. adv. We read all : 1. see Figure 3. entirely.7.1 All. Investigating the word “dou1” in the context of uttering a given sentence graph in Chinese. of several 2. any. each single 2. each and every Let us begin with giving the description of these four words as found.. The translations are as found in A Modern Chinese-English Dictionary [Ce. like for the fifteen different words in Chinese for the preposition “in”. everything 3. one by one. 6. quite any each every : 1. separately : every. In the next section we will first discuss these differences before dealing with the word “dou1”. each.g. each. However. One interesting remark is still to be made first. e.122 Chapter 6 every and each were considered to be expressible by the SKO-loop. all : all : arbitrary : 1. no matter which.

Hence the word graph consisting of one token and a single SKO-loop can be named ANY. So whereas “any” focuses on the single element. that can therefore be uttered as “any of (a) set”. however. This means that we can take the word graph SKO PAR ALI set for the word “every”. it comes forward in “regarded separately”. This can be interpreted as “something arbitrarily considered”. Here. In English we can say “For all x P(x)” and “For any x P(x)”. This last remark hints at a very important aspect. The association with “dui4” is “concentrating on”.Utterance paths 123 expressed in Chinese as “ren4yi4 x P(x)” or as “sou3you3 x P(x)”. Note the occurrence of the word graph for “any” in this graph. “every” focuses on the single element as part of a set. In Chinese the statement “dui4 ren4yi4 x P(x)” is slightly preferred over “dui4 sou3you3 x P(x)”. there is the extra remark that “every one of two or more things” is part of the meaning of “each”. The Chinese word corresponding to “for” is “dui4”. meaning that there may be more elements. “every” and “all”. The single element aspect comes forward in “every” with the meaning “each single”. A SKO-loop then can be read as “something informationally dependent only on itself”. In the meanings given for the Chinese words the single element aspect comes forward . In “each”. EQU 2 Note the occurrence of the word graph for “every” and the extension that can be uttered as “with cardinality greater than 2”. The entry for “every” also mentions “all possible”. namely that a single element of a set is focused upon. We take this as our starting point for finding word graphs for “each”. The SKO-link was introduced with meaning “informationally dependent on”. which is synonym with “every”. Now we can extend this graph to obtain the word graph for “each”: SKO PAR ALI PAR ALI set number ORD cardinality .

one of which is “all”. The third notion expresses the concept of “whole”. We . also in English the words “each” and “every” are not always distinguished very precisely. that can assume different values. Under the heading “Quantity” we found “unity”. as can be seen from phrases like “each and every”. but “each element of a set of one element” is in a subtle way less precise. “plurality” and “totality”. What is described in these expressions are instantiations of the totality. In a way the word “all” clearly differs from the three other words.124 Chapter 6 in “ge4ge4”. The focus in all three meanings is on the totality. However. via “per”. This brings us back to Kant’s ontology. this graph could also be given as the word graph for “suo3you3” in the form PAR ALI suo3you3 EQU totality . Consider expressions like “half the butter” or “almost all butter” or “hardly any butter”. For this reason we could consider the following graph: PAR ALI totality . This means that the use of the SKO-loop for describing the word “all” is put in doubt. The given entries suggest that the words “ge4ge4” and “mei3ge4” are closer in meaning than is expressed by the given word graphs. We are left with the word “all”. In language we see that we can speak of “all butter” as well as of “all dogs”. and in “mei3ge4”. There were three entries given. As “all butter” is expressed in Chinese as “suo3you3 de huang2you2” but not as “dou1 huang2you2”. The word “de” then corresponds to the PAR-link. Thus “totality” is seen as an attribute of something. In both cases the totality aspect prevails. In mathematics we can speak of “every element of a set of one element”. The first two express the element aspect and the set aspect. We then have the word graph all EQU PAR ALI totality . via the phrase “one by one”. as we have already seen before.

In principle the possible utterings of such a sentence graph give all possible occurrences of the word “dou1”.1 the totality either describes an aspect of a set or of an object.6.Utterance paths 125 will discuss “dou1” in the next section. it turns out that for certain words.6. 2.6. They investigate certain sentences and partition them into the class of grammatically correct and the class of grammatically incorrect sentences. Note that “dou1” cannot be used for quantifying mass words. This is an adword. However. because of the PAR-link. 6. In a paper of Cheng [Cheng. However. In both cases we do NOT have attachment to a verb. An essential remark in Cheng’s paper can be found in her example (2) that reads (she does not give numbers 1. in order to investigate the possibilities for uttering. Note that this approach differs methodologically from that used by Cheng and others. However. This presupposes that the type of word is known. Cheng comes to the conclusion that “dou1” is a “nonmovable adverb”.2 Uttering the word “dou1” As we will see in Section 6. In this chapter we focus on uttering a sentence graph.1 “dou1” was seen as having a word graph like PAR ALI totality dou1 . We therefore studied the sentence graphs of several of the sentences mentioned by Cheng. there is disagreement on the type of that word. But to what type of word does it attach? As we discussed in Section 6. In Chapter 5 we discussed a grammar derived from the syntactic word graphs.7 the various ways to utter a sentence graph are controlled by the rules of the grammar that is assumed to be valid for the language. This means that we choose to focus on sentence graphs containing the word “dou1”. in Section 6. Then a discussion is held about similarity of sentences in the sense that “dou1” can or cannot occur in the same way as other words do. 3 or 4) EQU . we do not have rules for uttering such a sentence graph yet. like the Chinese word “dou1”. 1995] on “dou1” quantification a discussion about the possible occurrence in a sentence can be found.

For this example sentence we would give the sentence graph tamen EQU CAU PAR .” (ASP. Any statement about “those students” in principle includes all of them. The word graph for “dou” has been indicated here completely. in principle all the people described are meant. is tamen EQU CAU PAR ALI ALI lai dou EQU totality . those students. have a plural interpretation. We now go through the example sentences. mathematically. they.1 of her paper. We may therefore consider the example sentences without the word “dou” to see whether the meaning of the sentence is essentially changed. In the following examples we will just link a black token to the word with plural interpretation. Before doing this let us remark that statements like “tamen”. We decided to focus on this aspect and investigate how various sentence patterns allowed the uttering of “dou” for the other example sentences in Section 2.126 Chapter 6 Tamen They dou all hen very xihuan like wo. apart from the tense. I She remarks that “dou(1)” here quantifies a noun phrase NP to its left and that the NP must have plural interpretation. But that means. dou lai-le (1) tamen they all come-ASP “They all come. that the NP must describe a set. The same holds for “neixie xuesheng”. is referring to the function of the word “le”) The sentence graph. ALI lai . for aspect. or “neixie xuesheng”. If the word “tamen” is used. If not the word “dou” is only used to express emphasis.

to express emphasis. Here both “tamen” and “neixie xuesheng” have plural interpretation. Again “dou” is used for emphasis. the only word with plural interpretation. So. (3) tamen they dou all xihuan like neixie those xuesheng student. The sentence graph is: tamen EQU CAU PAR CAU EQU neixie xuesheng ALI xihuan . but this is considered to be as good as simply “tamen xihuan neixie xuesheng”. There is no essential difference with example sentence (1) as the word “dou” quantifies “tamen”. we can say that “dou” can only be used. We now also consider this simple SVO-pattern in a sentence where both subject and object have plural interpretation. Example sentence 2 involves a transitive verb. namely quantifying the subject “tamen”. and therewith the black token in the sentence graph. It could have been left out here. Note that the sentence “tamen lai-le” has essentially the same meaning. This again shows that words with plural . (2) tamen they dou all hen very xihuan like wo I “They all like me very much”. Suppose we would like to say: “they like all those students”. In Chinese we might say “tamen xihuan suo you de neixie xuesheng”. However. for a subject with plural interpretation. as an utterance rule. and moreover it mainly functions as giving emphasis. The word “dou” could have been omitted. This speaks against the assumption that “dou” is an adverb. Yet the word “dou” can only be used once. it cannot be an adword for “neixie xuesheng”.Utterance paths 127 We see that “dou” is an adword of “tamen” and not an adword of “lai”.

128 Chapter 6 interpretation do not have to be combined with words expressing universal quantification. Replacing “dou” by “tamen” we have a similar sentence (5) zhexie these “These xuesheng student students I wo I like xihuan like them”. “them” is used for the reference. Already from the third line in which. (4) zhexie these “I xuesheng student like all of wo I dou all xihuan like students”. in the translation. the direct adword function with respect to “those students” is evident. the reference function is clear. First there is the pattern of combination of “dou” and negation. The word “zhexie” already describes the plurality aspect. Three other patterns will be considered. Cheng gives the sentence (6) neixie xuesheng wo those students I “All those students xiangxin believe I believe Lisi Lisi Lisi dou all likes xihuan xihuan them”. “dou” seems to have a reference function similar to that of determiners and pronouns. Next to the emphasis function. We consider the sentence graph in both cases. Let us consider the example sentence (7) neixie those ren √i person meiyou √j not kan-guo read-ASP neiben that shu. these Here “dou” is used as a reference word. Also. . book where “dou” can be inserted on two places. without attention to the tense aspects. tamen they where the only difference is that the referring word “tamen” now follows the verb. from the translation. Also in the case of an embedded sentence we encounter the reference function of “dou1”.

A sentence in SVO-form may be uttered in SOV-form. In the sentence: (8) b. (8) a. Finally there are the BA-pattern and the BEI-pattern. either around CAU EQU neiben shu . (j) Not all of those people read that book. neixie those xuesheng student dou all ba BA neiben shu mai-le that book sell-ASP “All those students sold that book. zhangsan zhangsan ba BA neixie those shu book dou all mai-le sell-ASP “Zhangsan sold all those books”. but then the auxiliary word BA has to be used. In the latter case in Chinese the subject “neixie ren” is taken out of the frame.” Here the subject has plural interpretation and “neixie xuesheng” can be followed by “dou”. kan This graph gives the basic structure without the NEG-frame: those people read that book. and again “dou” has a reference function. The NEG-frame can now be inserted in two ways. ALI kan or around the whole graph (j). as emphasis. . (i) . The meaning is then respectively (i) All of those people did not read that book.Utterance paths 129 neixie ren EQU CAU PAR CAU ALI EQU neiben shu .

” Again the uttering of “dou” is directly after “zhexie laoshi”. but uttering it here is not considered. neixie xiaohai dou bei Lisi qifu-guo bully-ASP those children all BEI Lisi “Those children were bullied by Lisi. However. and “dou” can occur only here. the essential meaning could also be uttered in “S V O”. Another. “dou” would have a reference function. The uttering of this sentence graph is “All S V all O”.” Here the uttering of “dou” is directly after “neixie xiaohai”. zhangsan zhangsan bei BEI zhexie these laoshi teacher dou all ma-le scold-ASP “Zhangsan has been scolded by all these teachers. . usually used for expressing the completed tense. because of the plural aspect of both S and O. often used. When uttered after “Lisi”. In Chinese we can say “S dou BA O V”. In case both S and O have plural interpretation. that has plural interpretation. ALI PAR V We have assumed that both S and O have a plural interpretation and that there are two adwords “dou” attached to them. This is the uttering rule for “dou” in case of a sentence with BA-pattern. Again Cheng gives two example sentences. pattern is the BEI-pattern. that has plural interpretation. In abstract sense we consider the following sentence graph: S EQU CAU PAR CAU EQU O . (9) b.130 Chapter 6 the object can be quantified by “dou”. and can occur only here. again only one can be emphasized by “dou”. by uttering “dou” after “neixie shu”. (9) a.

Sentence b is not correct. She gives example sentences (30) a. as emphasis.1 Cheng states that “dou” falls within the class of nonmovable adverbs.” b. In that case the word puts emphasis on the noun. Dou falls within the class of nonmovable adverbs like . It is therefore remarkable that in Cheng’s paper “dou” is considered to be an adverb. From the knowledge graph point of view we see that “dou” is considered to be an adword attached to a noun with plural interpretation. To put it in the words of Cheng. one as quantifier. We quote: “Dou is not a time adverb or attitude adverb.2. after “neixie xiaohai” or after “zhexie laoshi”. although that is not absolutely necessary. Zhangsan yijing hui jia-le Zhangsan already return home ASP “Zhangsan has already return home. In Section 3. like “yijing”.Utterance paths 131 Let us consider the abstract situation again. The other function is that of a reference word. we think that “dou” is a head taking a noun as complement. In neixie those xiaohai children bei BEI zhexie these laoshi teacher ma-le scold-ASP we can utter “dou”. “Yijing” cannot “move” from after the subject to the front of the sentence. It has two functions. jintian today wo I bu not shufu comfortable “Today I don’t feel well. * Yijing zhangsan hui jia-le. in which both S and O have plural interpretation. wo jingtian bu shufu we see that “jintian” can be moved to the front. In the example (20) a. already. and it cannot appear before the subject. It is thus not a movable object. In case both have plural interpretation both “O dou BEI S V” and “O BEI S dou V” are allowed. The BEI-pattern turns the sentence “S V O” into the sentence “O BEI S V”.” b. Uttering “dou” is now only possible in a sentence with one of S and O.

From the point of view of knowledge graph theory the difference should come forward in the sentence graph. .” This conclusion is mainly based on grounds concerning the distribution of adverbs in uttering. For “today” the graph is: ts . by definition. te] time interval . One is nonmovable and the other one is. So in both sentences we have the basic structure he EQU CAU PAR ALI return . let us consider the two time adverbs “already” and “today”. However. The act. before the uttering time ts of the ALI PAR verb v today EQU SUB ALI EQU [tb. has finished at te.132 Chapter 6 yijing ‘already’. of the two adverbs. described by the verb v. We consider the two sentences “he returned today” and “he returned already”. The adverbs attach to the verb. EQU adverb The difference must lie in the word graphs. For “already” we have: ALI PAR ORD EQU EQU verb v already EQU te sentence. the semantics.

6. although both graphs attach to the verb v. i. i. Although we are dealing with two time adverbs here. Let there be a set of utterings for a sentence graph. as “the structure is the meaning”. utterance path. A specific parse tree reflects an utterance path. The following figure schematically describes the relationships between parse tree. So there are as many parse trees for the one given sentence graph as there are utterings.Utterance paths 133 The difference between “already” and “today” is considerable. We have seen that a sentence can be seen to consist of chunks. in Chinese. It must have a parse tree based on the grammar assumed to be valid. like Cheng does. Each uttering should be grammatically correct. chunks and sentence graph..e. These chunks are on the one hand parts of the parse tree and on the other hand parts of the sentence graph. we now want to start a more general discussion on uttering a sentence graph. both words are adverbs. utterance path. There is a considerable semantic difference.. the structure of the sentence graph seems to have profound influence on uttering.7 Uttering and Grammar Having considered the uttering of very specific types of words. the way it occurs in the sentence graph does. PARSE TREE UTTERANCE PATH CHUNKS SENTENCE GRAPH Figure 6. This brings forward an important aspect of uttering. That “dou” has the distributional properties of a nonmovable adverb like “yijing” seems to be no reason to consider it to be a nonmovable adverb. when we focus on the explicit description of the time aspects.e. So also for “dou” we may expect that its meaning plays an important role.7 Relationships between parse tree. . chunks and sentence graph.

For this reason the subgraph can be uttered as “the dog”. 6. or word groups. The syntactic graph of a determiner in Figure 5.134 Chapter 6 Our goal is to describe how a sentence graph can be uttered in a grammatically correct way. Let us now consider the production rule 6. expressing that a noun can be replaced by a determiner followed by the noun. Whenever occurring in a sentence graph the allowed order can be indicated by an arc from the determiner to the noun.1 An introductory example We will choose.. Rule 6: N → det N gives us a possible uttering rule for a determiner attached to a noun. Let us. i. For this we have to indicate which words. again. as an example. the grammar for English derived in Chapter 5 from the way syntactic word graphs can combine. but not as “dog the”. it merely expresses the possible order of the words in juxtaposition. where the EQU-link is the word graph for “the”. Note that this arc has no label. This is coming forward from the syntactic graph. This typically is ruled by the syntax of a language and therefore the possible juxtaposition of two words is ruled by the production rules of the grammar.e. N → det N. which might therefore be called an “uttering graph”. a determiner and a noun. from the direction of the SKO-arc. phrases.7.6 is det ALI SKO ALI N . . In a sentence graph we might find the subgraph EQU ALI dog . may follow each other. consider the uttering of the words “the” and “dog”. Deleting the links of the sentence graph gives thus a graph containing only such “uttering arcs”. N → N det. The word “the” might be given the semantic word graph: the EQU EQU . The important aspect is that the rule does NOT state 6.

A complication comes in when rule 5 is applied twice: N → adj N → adj adj N. we want to investigate rule 5: 5.Utterance paths 135 Before systematically investigating the 18 production rules of our chosen grammar. In a sentence graph we might find the subgraph: poor ALI PAR PAR ALI small ALI Syntactically we have: adj ALI PAR dog PAR ALI ALI . then we can say “small dog” and NOT “dog small”. The part of the uttering graph corresponding to the considered subgraph is now poor ALI ALI small ALI dog . Let us consider the words “poor”. . N → adj N. adj N . Here too an ordering for uttering an adjective and a noun is forced. The possibility of generating a second adjective by rule 5 implies that for the two adjectives there should be uttering arcs from one adjective to the other. “small” and “dog”. Suppose we have the words “small” and “dog”. This implies that two adjectives can be uttered in juxtaposition before uttering the noun. The uttering arcs are corresponding to the PAR-arcs towards the noun. We may utter “poor small dog” as well as “small poor dog”.

but not the other way around. “the” must be uttered first. and let us give both the part of the sentence graph and its uttering graph. The two adjectives “poor” and “small” can be . All four words have to be uttered. This implies that there can be an uttering arc from the determiner to the adjective. “dog” must be uttered last. Rules 5 and 6 can be combined. As there are only incoming arcs for the noun. To finish this introductory example let us see how the semantic graph can be uttered. Let us now consider the phrase “the poor small dog”. The arcs from the determiner to the two adjectives stem from the combination possibility for rules 5 and 6. The semantic graph is EQU EQU PAR ALI the ALI poor ALI PAR small dog and its uttering graph is EQU . dog The arc from the determiner to the noun has been discussed. but only in the following order: N → det N → det adj N. As there are only outgoing arcs for the determiner in the uttering graph.136 Chapter 6 where we have maintained the typifying ALI-arcs for reasons of clarity. the poor ALI ALI small ALI .

7.3. 6. has 24 Hamilton paths and would look like ALI today Peter EQU ALI book ALI read . The uttering graph looks like ALI today ALI Peter EQU book ALI read and has one Hamilton path. For the imaginary language X. for the Serbocroatian grammar.Utterance paths 137 uttered between “the” and “dog” in both possible orderings. The uttering graph for the sentence graph given in Figure 6. Note that the two “uttering paths” in the uttering graph are Hamilton paths as all vertices are contained in the paths. We formed the following English grammar rules: . only uttering was possible.2.2 Uttering rules from production rules Let us recall the grammar described in Section 5. In fact the number of possible uttering paths is equal to the number of Hamilton paths in the uttering graph.

The rules 2. V → adv V 16. Given a sentence. Our reasoning will be restricted by this choice. AP → V N 11. 3. VP → V 12. V → V adv 17. V → V AP 18. 3 and 11 concern simple instantiation and therefore do not infer a condition on the order of uttering words or phrases. 7. 6. 9. S → NP VP NP → PN NP → N N →N N N → adj N N → det N AP → prep N N → num N N → PN N 10. Note that this is just one of many grammars that may be considered. 8. 5. adj → adv adj .7. 6.2.1 Rules involving word types only We start by considering the rules in which a noun N is involved. 4.138 Chapter 6 1. These rules are rules 4. This leaves 15 rules to be considered. 8. 6. VP → V N 13. 5. and no phrases. 9. It was already pointed out that there is a difference between the role of grammar rules in traditional parsing and the role they play in uttering a sentence graph. the problem is to find a parse tree and the . VP → V PN 14. V → V V 15. 2.

However. Remaining are rules 14. Let us now consider the phrase “The three mean tall men”. These arcs will be called uttering arcs. and 16 involving V. so the generation of this phrase should start with rule 6. which forces us to use rule 8. Their order is irrelevant. as e. both cannot be used. Then the numeral must be uttered. Note that there cannot be an u-arc from adv to N. N → N N generates a juxtaposition of two nouns. The schema describes the uttering rules for parts of the sentence in which the five rules must be used when parsing the uttered phrase. This now means that for uttering a part of a sentence graph in which a noun occurs. when applying rules in arbitrary order we might generate phrases or sentences that are not correct. by rule 5. The second noun cannot be taken to generate for example an adjective. So only certain orderings of rules are possible. Rule 4. For example we may say “he just came”. in which a PN is generated before the noun is similar to rule 6. Rules 15 and 16 describe that an adverb attached to a verb can be generated in both orderings. u-arcs. but not . However.Utterance paths 139 order of applying rules is to be determined. Rule 9. For the uttering ordering expressed in the schema. 15. An important other example is the combination of rule 5 and rule 6. in “very big”. that would then stand between the two nouns. We can say “severe thunder storm” but not “thunder severe storm”. We know that the determiner “the” has to be uttered first. there are uttering orderings that can be indicated as: det noun num adj adj . while from det and num there should be arcs to adv. Rule 18 puts an adverb in front of an adjective. Also note that rule 18 can be repeated as we can say “very very big”. given a sentence graph. We will use our own knowledge to decide on the possibility of utterings.g. as “the small dog” is a possible uttering. The two adjectives can now be generated by applying rule 5 twice. this means that adv should be added and with arcs to adj. but “small the dog” is not.

In a sentence graph an AP starting with a preposition can be determined by that preposition. There are two conclusions here. 12. in principle for uttering a verb-adverb attachment we need more information and cannot give a general uttering rule here. Rule 17 puts an AP behind the verb V.7. 7. 10. which in this case is just the word “go”. V → V V is particularly interesting because of what we discussed in Section 6.2 Rules involving phrases We are left with rules 1. The sentence “He must go” has sentence graph NEC EQU CAU ALI ALI he (person) go . 13 and 17.140 Chapter 6 “he came just”. We might therefore speak of prepositional phrase too. The first V of V V usually is an auxiliary verb. but prefer to use the more general notation AP. Note that the preposition.5. but this means that there should then be u-arcs from the preposition to all the word types that can be uttered before the noun in the noun phrase. but not “he hard worked”.2. This is just . Rule 14. This noun could be extended to a noun phrase. Let us begin with the rules that involve the adverbial phrase AP. The frame expresses the auxiliary verb “must”. was one of the chunk indicators in Chapter 5. We have to start uttering with the PN “he”. Then the content of the frame must be uttered. So. An uttering. the frame. as considered in Section 6. Rule 10 generates the order V N as an AP. but then immediately the auxiliary verb. Rule 7 shows that such a phrase may start with a preposition. while we can say “he worked hard”. 6. Rule 7: AP → prep N generates a preposition before noun. because English is an SVO-language. like “drinking coffee” may occur as an AP bringing two verbs in juxtaposition.7.1. which in the sentence graph is represented by a frame. must be uttered.2. a link word.

The 8 possible sentences are 1. determine the noun in the NP.3 Uttering paths for the extended example We have discussed our representation of the uttering graph. The tall mean man hit in the garden with a very big stick the small poor dog. The mean tall man hit in the garden with a very big stick the poor small dog. The tall mean man hit in the garden with a very big stick the poor small dog. then add a u-arc from the end of the noun phrase (which is the noun) to the verb. 4. The APs can be uttered in only one way but in two possible orderings. We delete all links between tokens but maintain the EQU-arcs and ALI-arcs in order to indicate the words. . The AP should be uttered after the verb has been uttered. 6. The mean tall man hit in the garden with a very big stick the small poor dog. Rule 17 restricts the position. Determine the verb and. Rule12: VP → V N shows that there should be a u-arc from the V to the noun N. 2. The mean tall man hit with a very big stick in the garden the poor small dog. in any order. in an uttering. of an AP.3. Repetition of rule 17 is possible and. 6. Finally. 3. The verb phrase starts with a verb. two APs can be uttered after each other. We will now apply our findings to the extended example of Section 6. as is determined by rules 12 and 13. rule 1 reflects the SVO-language and we have already discussed how to determine the NP.Utterance paths 141 a consequence of the fact that we took rule 10 into our grammar. The 2×2×2 = 8 possible utterings correspond to the 8 Hamilton paths of the uttering graph. These are the u-arcs consistent with the considered grammar. 5. by the incoming CAU-arc. For frame words like “with” and “in” we use auxiliary tokens to avoid too complicated graphs. and to all that can be generated before that noun. here.7. The tall mean man hit with a very big stick in the garden the poor small dog. There are two ways to utter “the mean tall man” and “the small poor dog” parts.

allowing the use of commas.8 Uttering graph for the extended example. 3.142 Chapter 6 7. but with interchange of the APs “in the garden” and “with a very big stick”. 7. The tall mean man hit with a very big stick in the garden the small poor dog. The reader may read off the given 8 sentences by following the 8 different Hamilton paths. This would again increase the number of possible utterings. 8. mean EQU EQU the tall EQU ALI man the EQU very EQU big EQU hit ALI ALI ALI ALI with ALI ALI stick EQU EQU in garden a small the poor EQU ALI dog Figure 6. Figure 6. 6. . Note that the APs could be uttered after “the poor small dog”. 2. The mean tall man hit with a very big stick in the garden the small poor dog.8 gives the uttering graph of the sentence graph. 4. Sentences 5. 8 are as 1.

1979. the specific notion of information extraction was relatively new in the series of Message Understanding Conferences (MUCs). Sager. As a core of language technology.1 Introduction Over the last decade there has been a growing interest in developing systems for information extraction (IE). although varying in every case. IE technology has not yet reached the market. IE systems represent the need and ability to manipulate and analyze information automatically. information extraction refers to any process that creates structured representation of selected information drawn from one or more texts. In a broad view. Usually this process involves the identification of instances of a class of events or relationships and the extraction of the relevant arguments or relationships in a natural language text. but it was thought to be of great . 1987]. The idea of reducing the information in a document to a tabular structure goes back to the early days of NLP applications [DeJong.Chapter 7 Information Extraction 7. by integrating a variety of natural language processing technologies. Schank & Abelson. An enormous amount of information exists only in natural language form. The output of the extraction process. is finally transformed to the content of some type of database. However. 1977.

IE has achieved notable success [MUC-3.). Information Extraction is a more limited task than “full text understanding”.144 Chapter 7 significance to information end-user industries of all kinds as. One could translate a document and then extract information from the result or change the order of these procedures. IE technology should not be confused with the more mature technology of Information Retrieval (IR) that selects a relevant subset of documents from a large volume set by query. MUC-6. arson. the process of an IE system includes two major parts: (1) Extraction of the individual “facts” from the text through local text analysis. i. as a more focused and well-defined task. Finally the pertinent facts are . finance companies. publishers and other document-dependent managing agencies. we expect the representation of all the information in a text in an explicit fashion. for example. In full text understanding. etc. MUC-7. For example. or as a form of database (with the fillers of the template slots corresponding to database fields). Moreover. Grishman & Sundheim. victims. an IE system would extract the date. perpetrators. we restrict to the semantic range of the output: the relations we will represent. It should be noted that IE is not a wholly isolated information technology.2 The State of the Art of IE Generally. MUC-5. Cowie & Lehnert. the filled template. targets and type of attack (bombing. in the domain of terrorism. MT (Machine Translation) and IE are just two ways of producing information in applications and can be combined in different ways. 1996. can be managed either as a compressed text itself. In the last decade. 1996].e. For example. MUC-4. and (2) Integration of these facts to generate new facts through inference. as the task given in the MUC-4 evaluation (1991). To be a trend of “understanding” by extracting information from texts. Any application of IE is usually preceded by an IR phase. 7. This means that the product of an IE system. the IE process is computationally less expensive than in-depth natural language processing. banks. and the allowable fillers for each slot of a relation. location. Since many phrases and even entire sentences can be ignored if they are not relevant to the domain. a simpler MT system might only be adequate to translate the contents of templates that resulted in an IE process. In information extraction. IE extracts information from the actual text of a document.

or specified by computational linguistics. set or activity. IE development relies on recent advances in empirical NLP techniques. or scenario. dates and monetary amounts. a scenario is specific to particular events or relations to be extracted. (5) Scenario Template filling requires the detection of relations between template elements as participants in a particular type of event.Information extraction 145 represented and translated into the required output format. such as word sense tagging. (1) Name Entity recognition requires the recognition and classification of named entities such as organizations. Currently the most successful systems use a finite automatic approach. certain artifacts. Practically. Recent research [Church et al. (4) Template Relation filling requires filling a two-slot template representing a binary relation with pointers to template elements standing in the relation. Five separate component tasks. and subtype. description as supplied in the text. 2000]. were specified by recent MUC evaluation (MUC-7). definite noun phrases and their antecedents and pronouns and their antecedents. with patterns being derived from training data and corpora. These include variant forms of name expression. Recent research on IE was stimulated in large part by the MUC evaluations. (3) Template Element filling requires the filling of small scale templates (slot-filler structures) for specified classes of entities in the text. and a template refers to the final tabular output format of the IE process. such as organizations. and the construction of an object-oriented structure recording the entities and various details of the relation [Humphreys. 1996] has also shown that a number of quite independent . Many relatively independent modules within some general knowledge-based AI program have achieved significant success for a range of linguistic tasks. sentence alignment and so on. locations. and their locations. which illustrate the main functional capabilities of current IE systems. syntactic parsing.. persons. According to the terminology established by the MUC. persons. (2) Coreference resolution requires the identification of expressions in the text that refer to the same object. with slots such as name (plus name variants).

without building a meaning representation of the overall text. a Text Zoner. 2000]: 1. There have been IE systems developed in groups stimulated by the MUC.. there have been attempts to derive a richer meaning representation of the text with less task. maximally filled template. Such approaches were motivated by the belief that high level of precision in the IE task will not be achieved without attempting a deeper understanding of at least parts of the text. FASTUS [Appelt et al. with richer meaning representation.146 Chapter 7 modules of analysis by learning and statistical methods can be built up independently from data. Many developers of IE systems have opted for robust shallow processing approaches that do not employ a general framework for “knowledge representation”. 1995]. 1995]. domain-specific heuristics for merging partially filled templates to yield a final. SRA [Krupka. which turns a text into a sequence of sentences. MITRE [Aberdeen et al. 1995].. the functionalities shared by most systems alternatively include the following [Cowie & Wilks. . 1995]. It was claimed that the MUC-6 evaluation showed that such an approach. 2. overall performed not worse than the shallow processing approaches [Cowie & Wilks. rather than coming from either intuition or some dependence on other parts of a linguistic theory. 1995] and the SRA and MITRE MUC-6 systems [Krupka. Aberdeen et al. and so on. which turns a text into a set of segments. A typical discussion about the NLP techniques used for IE is that of Hobbs [Hobbs.. word sense. 1993]. as exemplified in the systems of FASTUS [Appelt et al.. nor representing and using world and domain knowledge in a general way to help in resolving ambiguities of attachment. 1998)].. NYU Proteus system [Yangarber & Grishman.. quantifier scope. In other words. 1992]. coreference. 1995]. such as POETIC [Mellish et al. Such shallow approaches typically rely on collecting a large number of lexically triggered patterns for partial filling templates. LASIE [Gaizauskas et al. 1995]. such as the discourse model and intermediate representation used in the design of the LASIE system [Gaizauskas et al. a Preprocessor.and template-specific approaches. there may even be an attempt. UMASS [Fisher et al. 1995]. However. 2000].. 1995.. with alternative features to the IE task by applying NLP techniques. According to Hobbs’ paper.

Information extraction


3. a Filter, which turns a sequence of sentences into a smaller set of sentences by filtering out irrelevant ones. 4. a Preparser, which takes a sequence of lexical items and tries to identify reliably determinable small-scale structures. 5. a Parser, which takes a set of lexical items (words and phrases) and gives a set of parse-tree fragments as output. 6. a Fragment Combiner, which attempts to combine parse-tree or logical-form fragments into a structure of the same type for the whole sentence. 7. a Semantic Interpreter, which generates semantic structures or logical forms from parse-tree fragments. 8. a Lexical Disambiguator, which indexes lexical items to one and only one lexical sense, or can be viewed as reducing the ambiguity of the predicates in the logical form fragments. 9. a Coreference Resolver, which identifies different descriptions of the same entity in different parts of a text. 10. a Template Generator, which fills the IE templates from the semantic structures. It should be noted that there exist disputes on the practice of IE with the above organization. For example, module 8 could be performed early on lexical items, or later on semantic structures. Within the process under module 5, some people use a syntactic parser but the majority uses some form of corpus-derived finite-state patterns to represent the lexical sequences, which process would be called “semantic parsing” [Cowie et al., 1993]. For an IE task, defining templates is difficult, which involves the selection of the information elements required, and the definition of their relationships. The definition consists of two parts: a syntactic description of the structure of the template (often given in a standard form known as BNF-Backus Naur Form), and a written description of the rules on filling the templates and instructions on determining the content of the slots. The actual structure of the templates used has varied from the flat record structure of MUC-4 to a more complex object oriented definition used for


Chapter 7

MUC-5 and MUC-6. For example, a person object might contain name, title, age and an employer slot, which is a pointer to an organization object. Such newer object style templates make it easier to handle multiple entities which share one slot, as they group together the information related to each entity in the corresponding object. However, the readability in printed form suffers a lot, as much of it consists of pointers. Although many IE systems have been proved effective for information extraction on limited domains, there are difficulties in construction of a large number of domainspecific patterns. Manual creation of patterns is time consuming and error prone, even for a small application domain. In the following parts, a new IE approach is presented, which is using a domain-independent model described by knowledge graphs. This method is to Extract information by Knowledge Graphs (KGExtract) from the natural language texts.

7.3 Overview of the Approach
The development of language engineering applications, information extraction (IE) in particular, has demonstrated a need for the full range of NLP and AI techniques, from syntactic part-of-speech tagging through to knowledge representation and reasoning. The task of information extraction can be seen as a problem of semantic matching between a user-defined template and a piece of information written in natural language. To this purpose, the structural parsing oriented semantic processing will be applied to IE, and we will show that such a new IE technique has considerable advantages in comparison to traditional approaches to information extraction from the texts. For example, the method presented here, which is to encode the input pieces of information and the filled template into knowledge graphs, is a kind of graphic representation and it is domain-independent. With respect to the advantages of the approach, our main points are: • a simple (but semantically rigorous) model; • the possibility of semantic checks guided by the model; • a domain-independent representation; • an automatic pattern acquisition;

Information extraction


Let us now discuss an approach that is a process consisting of the following phases: Lexicon and Morphology. In the procedure of extracting information from NL texts, the precise duties of the lexical and morphological processing depend on the language that is being processed. For example, Chinese, without orthographically distinguished word boundaries, will require that some word segmentation procedure will be applied, but English can skip this procedure. The most important problem faced in this phase is to handle the proper names in a text. Because most extraction tasks require the recognition of persons, companies, government organizations, locations, etc. In addition to name identification, this phase must tag word types to words. We have discussed 8 word types in English in Chapter 5, such as noun, verb, adjective, pronoun, numeral, preposition, adverb and determiner. For each word type, we have created its syntactic word graph that represents the syntactic function of a word type. In Chinese we also chose 8 word types, but we replaced the “determiner” type by the “classifier” type. Semantic chunk tagging. The role of this phase is to split the different sentences of the text into semantic chunks according to the chunk indicators, such as pairs of commas and/or period signs, auxiliary verbs, reference words, prepositions, “jumps”, etc. Here no more details about chunk indicators will be repeated since we have discussed the problem in Chapter 5. Partial Structural Parsing. Some IE systems do not have any separate phase of syntactic analysis. Others attempt to build a complete parse of a sentence. Most systems fall in between, and build a series of parse fragments. In general, they only build structures about which they can be quite certain, either from syntactic or from semantic evidence. In our approach, a partial structural parsing method will be applied to every sentence of the input to build a series of semantic chunk knowledge graphs. We will then combine these knowledge graphs of chunks to derive the information that is to be extracted. The partial structural parsing can make the patterns, that will be mentioned in the next phase, to be created easily. Domain-independent pattern creation. In order to extract information from texts, we have to have patterns representing entities and events occurring in the texts. Many


Chapter 7

IE systems use a pattern-matching approach, but the set of patterns has to be created for each target task or target domain. If we are using a “pattern-matching” method, most work will probably be focused on the development of the set of patterns. However, for different domains changes will also be needed to the semantic hierarchy, to the set of inference rules, and to the rules for creating the output templates. To summarize, other methods are domain-dependent, but our approach is domainindependent. This is the most important improvement of our approach in comparison with other systems. Pattern Merging. This is the main part of our approach. The procedure of merging patterns is actually to integrate two semantic chunk graphs into a bigger one. It can be repeated until the number of semantic chunk graphs becomes 1. Template Generation. Once a text has been fully processed and a domainindependent representation has been derived, this representation can be used to generate template structures.

7.4 Description of KG-extraction
This section addresses the theoretical issues related to the design and use of such a method for information extraction. We present some basic principles, and we illustrate a preliminary proposal for a model developed according to such principles. Definition 7.1 A KG-extraction is the mapping of unstructured natural language texts onto predefined, structured representations, or templates, which, when filled, represent an extract of key information from the original text, with special regard to a model based on knowledge graph theory.

7.4.1 Partial structural parsing
It is very important to realize that the role of parsing in an information extraction system is not to perform full text understanding, but to perform parsing on the relevant parts of the text. The shallow parsing techniques tend to be imprecise, although efficient and transportable, whereas the full parsing approaches tend to be

each application of extraction will be related to a different scenario. the segment of the text is assigned a label. For more detail about regularizing.2 KG-Structure In order to extract the information from text. that is the mapping of a sentence on a semantic sentence graph. we have to have patterns representing entities of interest in the application domain (e. It is performed along almost the same phases as structural parsing. which combines the various bigger semantic chunk graphs into a sentence graph. Our approach is to shift from a full structural parsing (which has been described in Chapter 5) to partial structural parsing. company takeovers.4. For the example of “executive succession”.. In many other current extraction systems most of the text analysis is performed by matching text against a set of patterns.Information extraction 151 very precise but not robust and efficient. management successions) and relations between such entities in the texts. there will be the following pattern: <company> forms joint venture with <company>. These patterns are domain specific.2 An example of representing patterns with knowledge graphs: 7. In another text about “joint venture”. The goal of partial structural parsing is creating the scenario patterns of information to be extracted. Most . there will be such patterns as: <person> retires as <position>. Definition 7.g. In general.2. 7.4. The partial structural parsing has the role not only of identifying syntactic structure but also of making the extracted information syntax independent “regularizing” or “standardizing” by constructing the semantic chunk graph. If the pattern matches a segment of the text. not obtaining the full sentence graph.4. <person> is succeeded by <person>. and one or possibly more associated features. we refer to the example in Section 7.2 Partial structural parsing is the mapping of a sentence that is in the input text onto a set of semantic chunk graphs of this sentence. but for the last phase of structural parsing.

a semantic chunk graph). In particular. reduced relatives. they have the same meaning. These patterns are varying as the system turns from domain to domain.. which is domain-independent... This regularization simplifies the scenario pattern creation.. which are manufactured by GM . The approach described in this section is to represent such patterns by knowledge graphs. 7.. etc.. GM.4. which manufactures cars . This provides a graphical representation of patterns (i. relative clause.e. person.152 Chapter 7 work will probably be focused on the development of the set of patterns.. location. Cars are to be manufactured by GM. . . GM is a car manufacturer. are mapped onto essentially the same semantic structure (i... GM is to manufacture cars. aims at easing the burden of pattern creation. For example. The KG-Structure. and position) are very important in the information extraction task. cars. cars manufactured by GM . etc. Although these various clauses have different syntactic structure.. because it is difficult and inconvenient to operate directly on the patterns. One can then operate conveniently on the knowledge graphs..e..3 Named entity recognition Named entities (like organization. we do not need separate patterns for Cars that are manufactured by GM. .. which is called KG-Structure. such as active and passive forms. It is necessary to recognize the proper names in the .. different clause forms. knowledge graphs). This is why there is only one pattern represented by a KG-Structure that is to be constructed as follows.

we recognize “San Francisco” as a place name. One of the strengths of this approach is that the KG-structure. This method does not require humans to design complicated patterns. Templates are then filled through matching a KG-definition with a KGstructure. which are represented with a KG-structure. The source KG-structures of the tagged sentence are integrated on basis of their similarity. 7.4. The basic idea involves the following: • • • • • Semantic chunk taggings are performed for each sentence in the text. its representation by a KG-structure is as follows: San Francisco EQU ALI place name .Information extraction 153 text and express them with a corresponding KG-structure. we will propose an almost automatic method to acquire patterns useful for IE from the text input. being domainindependent. and we do not list them one by one. supports the generation of the templates or summaries in a language different from that of the input texts. Inferencing and merging are performed by knowledge graph operation. Semantic chunks are extracted to be the source of the patterns. based on partial structural parsing. At the initial stage. Example In the sentence “I want to go to San Francisco”.4 Automatic pattern acquisition In this section. Other named entities can be represented with a similar structure. each tagged sentence is regarded to be a pattern consisting only of semantic chunks and named entities. The merging can be done by combining the most similar pair of patterns into one pattern. .

5 Inference and merging In many situations.154 Chapter 7 There are several methods to define similarity of structures that can be found in the literature.4. a slot in the template will be filled up.7 A worked out example We will extract the information from the following text: “George Grorrick. 7. and the task will be well-specified by a limited “domain model” rather than a full unrestricted “world model”. some of the information is only implicit. The KG-definitions of the templates are pre-defined and stored in the lexicon.4. In this chapter we will not use any specific similarity measure. partial information about an event may be spread over several sentences. of those aspects of it required for the IE task. which forms the KGdefinition of the template. this information needs to be combined before a template can be generated. because it will not involve finding solutions to all the problems of such a KG-structure. Once a text has been fully processed and a KG-structure pattern.3 A KG-definition is a knowledge graph that expresses the semantics of a template relevant to the information extracted. Once a KG-definition of a template matches with a KG-structure of a pattern. has been created. president of the famous hotdog . These template structures will include pointers to the entries (each entry is a knowledge graph expression of a location slot. Definition 7. 40 years old.4. this pattern can be used to fill template structures. In other cases.) in the lexicon.6 Generating templates An IE system does not require full generation capabilities from the intermediate representation (the knowledge graph). and needs to be made explicit through an inference process. 7. 7. This makes a generation feasible for IE.

we will fill each slot. The set of knowledge graphs corresponding to each slot is the following: EVENT PERSON AGE ALI ALI ALI EQU EQU EQU . was appointed CEO of Lafarge Corporation. and the output. He will be succeeded by Mr. one of the leading construction material companies in North America. As a result.” Input slots that stand for the information that people want to extract are: EVENT PERSON AGE OLD POSITION NEW POSITION NEW COMPANY LOCATION.Information extraction 155 manufacturer Hupplewhite. that should be obtained. John. is shown as: EVENT PERSON AGE OLD POSITION OLD COMPANY NEW POSITION NEW COMPANY LOCATION appointment George Grorrick 40 president Hupplewhite CEO Lafarge Corporation North America .

. which is the named entity (like organization. person. such as “CEO” and “George Grorrick”. Easiest is Indicator 0. • Indicator 4: names and numbers. and we chunk each sentence step by step according to chunk indicators. There are two sentences in our text. • Indicator 1: auxiliary verbs. Totally. as well as “by”. Besides them. John 7 : effective October 1. such as “of”. We get the chunks: 1 : George Grorrick 2 : 40 years old 3 : president of the famous hotdog manufacturer Hupplewhite 4 : was appointed CEO of Lafarge Corporation 5 : one of the leading construction material companies in North America 6 : He will be succeeded by Mr. which have been mentioned in Chapter 5. • Indicator 3: prepositions. There are 4 types of indicator. position). • Indicator 2: reference words. that are used in this example.156 Chapter 7 OLD POSITION OLD COMPANY NEW POSITION NEW COMPANY LOCATION ALI ALI ALI ALI ALI EQU EQU EQU EQU EQU . such as “will” in the second sentence. such as “he” and “one”. because they are very important in the information extraction task. “in”. location. we introduce a new indicator. we obtain the following 5 indicators: • Indicator 0: comma or period signs.

Indicator 2 concerns “one” and “He”. Note that “was appointed” has an auxiliary verb. John. So far we got: 1 2 : George Grorrick : 40 years old 31 : president 32 : of the famous hotdog manufacturer Hupplewhite 41 : was appointed CEO . 1 by October 1. The prepositions cut the sentence just before the preposition. Indicator 3 is easy too. 1 was abbreviated. but the sentence was already cut before “was” by the comma indicator. Just try to speak the sentence with natural pauses to see why we did this. cut before the form. like prepositions. 2 and 7. We now find: 31 : president 32 : of the famous hotdog manufacturer Hupplewhite 41 : was appointed CEO 42 : of Lafarge Corporation 51 : one 52 : of the leading construction material companies 53 : in North America 61 : He will be succeeded 62 : by Mr. next to chunks 1. So 611 : He 612 : will be succeeded.Information extraction 157 Note: Oct. Indicator 1 is about auxiliary verb forms and these. The computer can replace Oct. but these chunks already stand alone in chunk 51 and chunk 611.

Suppose some lexicon gives: president: officer of a company.158 Chapter 7 42 : of Lafarge Corporation 51 : one 52 : of the leading construction material companies 53 : in North America 611 : He 612 : will be succeeded 62 : by Mr. then we must have a huge list of “values” for this slot and if “president” is not on that list. to be followed by a computer. But a computer cannot make the jumps we make when we say “Now we make the semantic chunk graphs”. Only when the computer knows that POSITION and OFFICER are similar. The computer must know that and must know that CEO is short for Chief Executive Officer. In particular “of”. We have to give very detailed instructions to go ahead in the information extraction process. That is why it might be easier to generate names of slots ourselves. and also for CEO we would do that. We see CEO (Chief Executive Officer) as name. so “president” is a word that can fill the slot POSITION. then we would generate the slot POSITION for “president” and OFFICER for “CEO”. indicating a relational template “in North America” is a chunk with one slot filled in. then we would introduce the slot OFFICER. but it is the name of a position and so is “president”. Assume that the computer knows all the prescribed slot names for the candidate words. the computer cannot make semantic chunk graph 3. say POSITION. “Officer” and “president” are both positions. We do not chunk up further in view of what we did so far. Now back to the names and numbers. If the lexicon gives: president: position in a company. When we prescribe/give slots like POSITION. John 7 : effective October 1. “in” and “by” are linking two slots. So this has its disadvantages too. This is precisely the difficulty in artificial intelligence. Because we prescribe the slots the lexicon of the computer may determine which words can fill one of the slots . there is the possibility of reduction to just one slot. We want an automatic extraction procedure. which may make it easier to find out the value of the other slot.

What kind of names are they? The computer might find: George Grorrick Hupplewhite Lafarge : name of PERSON : name of PERSON or name of COMPANY or name of LOCATION : name of PERSON or name of COMPANY or name of LOCATION : name of PERSON or name of COMPANY North America : name of CONTINENT John October : name of PERSON : name of MONTH . The computer may know for example: EVENT PERSON AGE : appointment. October and the numbers 40 and 1 are supposed to be recognized as NAME and NUMBER respectively.Information extraction 159 EVENT. There is one EVENT per sentence. PERSON. CEO COMPANY : manufacturer. company LOCATION : So for these words an interpretation as slot fillers is assumed to be directly possible. Mr. because the lexicon says something like “have age” for “old”) POSITION : president. so that has been settled (as only thing so far). corporation. as these are to be described by nouns. succession (recognition of verb forms. Hupplewhite. Grorrick. but these are not among the given slots. POSITION. John. North America. The computer may look up “appointment” and “succession” as values of EVENT. Lafarge. AGE.g. leading is not passing this test) : He. LOCATION. : old (e. What other preliminary action can the computer take? The names like George. COMPANY.

| PERSON: John 7 : MONTH: October | NUMBER: 1. “leading” and “effective”. The computer should know that it has to conclude PERSON: George Grorrick. given the slots. Why? We are. the computer has to and here is where the reasoning gets tougher for getting the semantic chunk graphs. That is. CHUNK 2 NUMBER: 40 and SET: years stand consecutive.COMPANY or LOCATION: Hupplewhite 41 : EVENT: appointment | POSITION: CEO 42 : of PERSON or COMPANY: Lafarge | COMPANY: corporation 51 : one 52 : of the construction material | COMPANY: companies 53 : in | CONTINENT: North America 611 : PERSON: He 612 : EVENT: succession 62 : by | PERSON: Mr. From this we have to extract the desired information. COMPANY or LOCATION. only interested in nouns.160 Chapter 7 40 1 : value of NUMBER : value of NUMBER. We can also replace by slot names where possible. CHUNK 1 There are two names consecutive. This is . let us remove adjectives: “famous”. one for a PERSON and one for PERSON. Having done all this preparation we now have the following: 1 2 : PERSON: George | PERSON. COMPANY or LOCATION: Grorrick : NUMBER: 40 | years | AGE: old 31 : POSITION: president 32 : of the hotdog | COMPANY: manufacturer | PERSON. and more in particular in names and values. so “40 years”. Before going over to the artificial intelligence part.

CHUNK 612 EVENT: succession. but that means that this is the OLD COMPANY: Hupplewhite. This reference word still has to be dealt with. John. as the other noun occurring in this chunk is of type COMPANY: manufacturer. so “40 years” must be the value of that measure. CHUNK 7 MONTH: October is a TIME-concept which is not one of the slots. As “Mr. See later. it must be COMPANY: Lafarge. PERSON: John. and Hupplewhite is the name of a PERSON. CHUNK 51 one. CHUNK 31 POSITION: president. Conclusion AGE: 40 years. CHUNK 62 by PERSON: Mr. The place in the sentence suggests reference to COMPANY: Lafarge. There is no problem here. So the computer should forget about this chunk. Hupplewhite and Lafarge turned out to be companies. if necessary.” is not a name the computer should combine to: by PERSON: John or Mr. CHUNK 32 hotdog is FOOD. CHUNK 42 of COMPANY or PERSON: Lafarge COMPANY: corporation. CHUNK 41 EVENT: appointment POSITION: CEO. CHUNK 53 in CONTINENT: North America. CHUNK 611 PERSON: He. A position is attributed to a person and “president” follows “George Grorrick”. As output we so far have for sentence 1: . The only person is George Grorrick. OLD or NEW POSITION? The computer must choose OLD POSITION because of the place in the sentence. So we extract COMPANY: Hupplewhite. Here the main problem arises. See later. so LOCATION: North America is found.Information extraction 161 followed by AGE: old which has a measure. The link implied by “of” is to president. CHUNK 52 of the construction material COMPANY: companies. COMPANY or LOCATION. Expansion of CONTINENT gives LOCATION. The reference must be to a person mentioned in the first sentence. This chunk does not contain information relevant to the given slots. so OLD POSITION: president.

This must therefore. 32 and 41 partly. Succession Age and location are not mentioned at all in sentence 2. The computer has to know what appoint and succeed mean. The only position mentioned in chunk 41 is CEO.2. The chunks 41 and 612 are the vital ones. but COMPANY and PERSON and POSITION do occur. We used the chunks 1. The lexicon might give “appoint” = “give POSITION to”. be the NEW POSITION. For sentence 2 we so far have: EVENT PERSON AGE OLD POSITION OLD COMPANY NEW POSITION NEW COMPANY LOCATION . 31.162 Chapter 7 EVENT PERSON AGE OLD POSITION OLD COMPANY NEW POSITION NEW COMPANY LOCATION Appointment George Grorrick 40 years president Hupplewhite . From chunk 42 then follows that Lafarge is the NEW COMPANY and . This has to be decided by solving the OLD/NEW problem. implied by “give”.

George Grorrick 2. : Preposition. : Measure of time interval. This is implied by “gets”. For OLD POSITION and OLD COMPANY nothing is found. PERSON: John gets the NEW POSITION. used for describing a property. The information included in the lexicon might be as follows: 1. . : Number. so it is president and of NEW COMPANY: Hupplewhite. (2) First officer of a company. 40 year old 3. : Determiner. : Having fame. : No information in the lexicon. The lexicon might give “succeed” = “get POSITION of”. who is George Grorrick. To describe the event we choose to give nouns derived from the verbs used in the sentences. Thus we obtain “appointment” from “appointed” and “succession” from “succeeded”. For filling the other slots we should now discuss the role of semantic chunk graphs. The event is given by the whole text. The preposition “by” leads to the proper choice. We discuss the construction of these word graphs from an imaginary lexicon.4.8 Chunk graphs for the example There is a difference between the status of the slot “EVENT” and the status of the other slots like “PERSON”. We will describe four phases illustrating the discussion given sofar. 7. president of the famous : Name of a male person. The first phase gives the word graphs of the words occurring in the two sentences. as these form the essential parts of structural parsing.Information extraction 163 the first template is filled after filling in the location: “North America”. The position is that of “He”. part or attribute. (1) Leader of a state. : Of high age. “AGE”. etc.

: Preposition. : Preposition. : Synonym of “firm”. : Auxiliary verb. : (1) Building (2) The act of building. part or attribute. : Form of the verb “be”. used for describing that something is part of something else. : Kind of a company. . Hupplewhite 4. was appoint CEO of Lafarge corporation 5. one of the leading construction material company In : No information in the lexicon. used to express a situation. (2) Pronoun. used for describing a property. manufacturer : (1) Factory. used for describing a property. referring to a male person. part or attribute. used for describing an actor or a cause of a verb. : Form of the auxiliary verb “will”. : Give a position to. : Matter.164 Chapter 7 hotdog : Kind of sausage. he will be succeed by : Pronoun. : (1) Number. : Preposition. : Adjective. : Shorthand for Chief Executive Officer. built from the verb “lead”. : Determiner. : No information in the lexicon. 6. North America : Name of a continent. Get the position of. used to express acts in the future. referring to an element of a set. (2) Kind of company. : Preposition.

George: male Grorrick: 2. president: (2) leader officer first ALI ALI FPAR FPAR PAR state company EQU ALI rank . : (1) Causing effect. The word graphs for these words can now be constructed name 1. : Name of a month. John 7. 40: number ALI PAR ALI ALI EQU ALI PAR ALI EQU PAR ALI George person 40 number year: time interval age old: measure ALI PAR PAR ALI EQU high ALI ALI (1) 3. (2) Starting. effective October 1 : Address form for a male person.Information extraction 165 Mr. : Number. : Name of a male person.

ALI FPAR ALI CAU manufacturer company Hupplewhite: be 4. was: ts EQU ORD PAR EQU tb ORD CAU CAU ALI appoint: ALI position give . (3) PAR EQU PAR fame ALI hotdog: hotdog ALI FPAR ALI sausage CAU (1) manufacturer: ALI CAU factory (2) .166 Chapter 7 of: the: famous: (1) FPAR . (2) SUB .

(2) CAU CAU ALI construction: build material: company: company ALI ALI matter EQU ALI firm . (2) EQU SUB . (3) (1) (2) 5. element ALI EQU PAR set ALI of: the: leading: (1) FPAR .Information extraction 167 CEO CEO: chief FPAR ALI EQU ALI PAR ALI officer executive PAR ALI PAR of: Lafarge: (1) . one: number ALI EQU 1 . (2) SUB . (3) PAR leading (1) ALI ALI PAR building .

168 Chapter 7 in: ALI SUB name North America: PAR EQU ALI continent North America ALI EQU ALI person 6. he: male CAU CAU ALI will: ts EQU ORD act PAR EQU tb be be: ORD succeed: CAU CAU ALI ALI position get by: CAU .

If possible. 62. The information that is to be extracted may be found from the chunks 1. The second phase is to build chunk graphs from these word graphs. we want to avoid complete structural parsing. 612. “Position” and “Company” occur here and there. . These word graphs contain only little relevant information. 2. There are two persons: “George” and “ John”. Only if necessary. 53. but not specified of whom. we combine these chunk graphs into graphs for larger chunks. 32. 51. 7. “Age” is mentioned in “old”. effective: (2) tb CAU EQU ALI ALI PAR effect time name October: ALI PAR EQU ALI month October 1: number ALI EQU 1 . 31. : ALI PAR EQU ALI PAR ALI ALI PAR ALI person male person male Mr. Note that we use partial structural parsing.Information extraction 169 address Mr. name John: ALI PAR EQU John (1) 7. 52. 42. 611. Chunk 1 Chunk 2 : We only have at our disposal the word graph for “George”. also without specification. 41. : The three word graphs cannot yet be combined.

Chunk 41 : be CAU ORD CAU ALI sausage ALI ALI appointment give EQU ORD position PAR EQU ts tb CEO chief Chunk 42 : PAR ALI EQU ALI PAR ALI officer executive . ALI PAR ALI FPAR ALI company corporation . . we obtained ALI PAR ALI FPAR CAU fame company EQU ALI EQU manufacturer ALI hotdog . We choose alternative (2). We cannot introduce “Hupplewhite” yet.170 Chapter 7 Chunk 31 : As this chunk has only one word. the chunk graph is just the word graph for “president”. choosing alternative (1) for “of”. Chunk 32 : We choose alternative (2) for “manufacturer”. We cannot introduce “Lafarge” yet. Using the same methods as in Chapter 5.

Note that “Mr. Chunk 53 : SUB ALI EQU continent . Note that the “act” is “be succeeded” and that this verb was already processed to fill the slot EVENT. the word graphs cannot be combined.” and “John” can be combined if we assume that the fact that both word graphs contain the subgraph male justifies this. Chunk 62 : ALI PAR male person . North America Chunk 611 : This chunk has only one word again.Information extraction 171 Chunk 51 : As this chunk has only one word the chunk graph is just the word graph for “one”. ALI PAR ALI person . address ALI EQU PAR ALI CAU Mr. Chunk 52 : Without background knowledge. Compare with the act “was appointed” in the first sentence. Chunk 612 : succession EQU CAU ALI CAU act EQU ORD PAR EQU ts tb .

John” as filler of the slot PERSON. Remarks: Due to the fact that various names did not have a word graph. we will now discuss the construction of chunk graphs in detail. ALI John . To illustrate how important the expansion process is for obtaining our extraction goal. Yet it has to be represented in relation with “George” as both words belong to the same chunk. Chunk 1 : The word “Grorrick” was not encountered in the lexicon. this is rather unlikely. the filling of slots is still not very well possible. John . Likewise the word graph for “continent” may not contain the concept “location”. This is what makes it plausible that “Grorrick” is . Persons have both a first name and a family name. where we now wrote person in capitals as this is one of the slots. Chunk 7 : There is no possibility to combine the three word graphs.172 Chapter 7 This is an example of similarity of two word graphs.” and “John” are combined. It is a name in English. we may replace the graph by PERSON ALI EQU Mr. without mentioning “continent”. how much background knowledge is needed. What we need is relevant background information about “George”. Then the chunk graph for “by Mr. This holds both for the names of the slots and for the words occurring in the chunk graphs. Only Chunk 62 gives information when “Mr. However. The third phase introduces reasoning by expansion of concepts. John” is address PERSON name ALI EQU PAR ALI CAU PAR EQU Mr. Describing a continent will involve mentioning its location. in fact it is a first name. From the chunk graph we now read off “Mr. As an example we consider the slot LOCATION and the word “continent” in Chunk 53. Any word graph for LOCATION may contain several instantiations or associations.

Information extraction


a family name. This information should be available to the computer. Note that it might be possible for the computer to expand the concept “name” to obtain this information. If not, the computer has no way to handle the word “Grorrick”. The chunk graph becomes: name male

George PERSON Grorrick ,

family name


and we have found the filler for the slot PERSON in the first sentence. Chunk 2 : The relevant background information in this case is that “old” says something about a time interval. “40” stands before years (plural) and therefore relevant background information is that “years” is a set. If expansion shows that “40” can be the value of the cardinality of a set we can combine “40” and “years”. Expansion of “age” in the word graph of “old” may yield that it is a time interval. Now we can combine into: measure AGE

high time interval number









. cardinality measure year years This rather complicated chunk graph contains AGE. The filler of AGE may be chosen from this graph by noting that the words “40” and “years” occur in the text. Other words are due to the construction of the word graphs (like “high”) or due to the expansion process (like “cardinality”).


Chapter 7

Chunk 3 : The subchunks 31 and 32 each pose a special problem. In chunk 31 only “president” is mentioned. The word graph contains the concept “officer”. The slots OLD POSITION and NEW POSITION contain POSITION and a list of possible positions might not include “president” but may include “officer”. On the other hand expansion of “president”, by expansion of “officer”, may lead to the conclusion that “president” is a position. In both ways the link between POSITION and “president” can be established. What remains is the problem with OLD and NEW, as we already discussed just before we considered building chunk graphs. Solving that problem involves using the given text and not just expanding words of the chunk graph. In chunk 32 the word “Hupplewhite” poses the problem. Being a word in the middle of a sentence beginning with the capital H suggests that “Hupplewhite” is a name. This also uses the given text. Therefore we should, in principle, not process this word in this third phase. However, we will discuss it here. The fact that the word follows the word “manufacturer” implies that it is the name of that “manufacturer”. The chunk graph for 32, constructed sofar ties the name up with COMPANY, and therefore we have found another potential filler. However, also COMPANY only occurs in the slot names OLD COMPANY and NEW COMPANY, so that we have the same problem as for OLD POSITION and NEW POSITION again. Chunk 41 : The two subchunks can be combined due to the fact that expansion of “officer” gives that it is a “position”. From the combined graph we read off that CEO is a filler of POSITION. Chunk 42 : “Lafarge”, like “Hupplewhite”, must be a name and stands right before “corporation”, and as “corporation” is of type COMPANY we find another filler of OLD COMPANY or NEW COMPANY. The chunk graph looks like

Information extraction



corporation name







The two chunk graphs could be combined by remarking that, in chunk graph 42, the PAR-link, that represents “of”, has a token that should occur in chunk graph 41. The word order suggests that this is “CEO”. For the extraction of knowledge, in the form of slot fillers, this combining is not absolutely necessary. Note that the subchunks 41 and 42 already gave the answer. Chunk 5 : Chunk 51 must be interpreted as a pronoun, because “one” is used and not “1”, so we have to choose word graph (2). Chunk 52 poses the main problem, coming from the phrase “leading”, as an adjective may be combined with “construction” as a noun. However it is to be combined with “companies”. How can a computer interpret the three consecutive nouns “construction”, “material” and “companies”? The basic idea is to use expansion of the, small, word graphs given. Suppose we consider: Construction: (1) (2)



build Material:

matter company




We have to find proper expansion. Let us start by saying that a “company”


Chapter 7

does something, i. e., there is a CAU-arc going out from its token. This suggests that for construction we use the second word graph and then we can already construct company






The word “material” or “matter”, because of its standing on the right of “construction”, must be expanded to link up with “building” as an instrument. However, it can also be linked with “companies” if we expand “companies” as entities producing something. This would lead to a graph like company




Note now that without the word “material” we would read “construction companies” and the first linking of graphs would be the only one. The sentence might have had the phrase “house construction companies”. That phrase indicates that the companies construct houses. The computer has to know how to deal with a sequence of nouns. We might instruct it in the sense that the last noun is the essential one. This would then mean that the adjective “leading” is to be attached to “companies”. Then the relation with the forelast noun should be established. A “material company” asks for an interpretation of “material”. Is this a noun or an adjective? The lexicon only gave the noun interpretation. But then we can link by the expansion that companies produce, leading to the second graph. If “construction” is the forelast noun the first graph would result: the company constructs. The first noun in “house construction company” would be interpreted as the unspecified token, which would lead to the, correct, graph

so material used. . Finding a filler for this slot is enough to conclude that we have found the required filler as chunk 53 belongs to the first sentence. build We have given this discussion. We could use both word graphs for construction semantically. because of the interesting problem of constructing a chunk graph here. build The first noun in “construction material company” has to be linked to the graph constructed sofar for “material company”. during the building process. “material” is searched for. “Building material” can be both material of which a building consists (after the building) and material used for building (during the building). Chunk 53 yields a slot filler as the expansion of “continent” may lead to the information that it is a LOCATION. As a result. and. basically. For the goal of information extraction it does not give any answer in the form of a slot filler. there is very subtle ambiguity here. again. linking with the forelast word. the location of “Lafarge corporation”. of which LOCATION was one. we have company leading ALI CAU CAU ALI ALI PAR material ALI PAR produce CAU CAU ALI . Let us recall that for the EVENT “appointment” slot names were specified. More detailed expansion can lead to the information that it is the location of “companies” and. we would prefer to use the second interpretation. So. In practice. This is an example of the usefulness of partial structural parsing. But such a detailed analysis is not necessary. via “one”.Information extraction 177 company ALI CAU CAU ALI ALI house . as instrument.

which is the pronoun “He”. They are found by the fact that the word “succeed” is interpreted as “get the . The second sentence sofar has only led to fillers for the slots EVENT and POSITION. The sentence is rather short indeed. But this pronoun refers to a person. John” we can only process chunk 611. in principle enough information was found to fill the slots OLD POSITION. So we can refrain from processing this chunk. We have not been able to fill all the slots of the two templates corresponding to the two sentences.4. Chunk 7 : Chunk 7 only contains data referring to TIME. mentioned in the first sentence. In the fourth phase we do not expand word graphs with lexical information. This completes the first template. For the first template. However. John”. now the context information is used to decide upon fillers. “appointment”. but.178 Chapter 7 Chunk 6 : We now have to investigate the second template and find fillers for the slots. For the second template OLD POSITION and OLD COMPANY cannot be determined. The implications of this cannot be found by expansion within Chunk 6. So from the sentence part “He will be succeeded by Mr. Next to that “Mr. it was still to be decided which name should fill which slot. John” was localized as a PERSON. We have given a reasoning at the end of Section 7.7. to find OLD POSITION NEW POSITION OLD COMPANY NEW COMPANY president CEO Hupplewhite Lafarge . NEW POSITION. The pronoun “He” plays the vital role in determining the fillers for the slots NEW POSITION and NEW COMPANY of “Mr. We definitely need some extra reasoning. George Grorrick. but will only mention what can be decided in this phase for this example. as remarked before. We will not do this in detail. OLD COMPANY and NEW COMPANY. We already used the word “succeeded” to fill the slot EVENT with “succession”. but this was not chosen to be a slot name.

4.Information extraction 179 position of”. Therefore we get NEW POSITION: president and NEW COMPANY: Hupplewhite. that each can provide fillers for the chosen slots. 7. this company might be a company in South America. 2. who is George Grorrick. based on the types . on the basis of KGExtract. 3. 1. Clearly our parsing by chunks turns a text into a set of segments in much more detail. Due to the pronoun “He”. A Preprocessor. the important phases are: • • The third phase. the construction of chunk graphs. The slot LOCATION is still to be filled. The second phase. where the free token in the pronoun “He” is identified with the only person mentioned in the first sentence. All four phases should have their place in any automatic information extraction procedure. in principle formed by both sentences. We consider the whole text without filtering. Filters. • • The first phase. we can only conclude that the succession took place at the manufacturer Hupplewhite. Also this is covered by parsing by chunks. hardly gave any filler. However. For the “appointment” a location is mentioned. but the LOCATION slot of the “succession” has to remain open. we see that there are four phases. The fourth phase. A Text Zoner. Concluding. in which expansion of word graphs gave the opportunity to link potential fillers to slots. gave some possibility to attach names to slots. turned out to be of vital importance to decide on the proper choice of fillers.9 Discussion Let us consider the 10 functionalities mentioned by Hobbs. in which context information was used. referring to George Grorrick. However. A Filter. just the construction of word graphs.

5. See the discussion about Chunk 611. The formulation of Hobbs stresses the traditional representation forms of parse-tree or logical-form fragment. The traditional approach is to start with syntactic aspects. A Template Generator. 8. Let us end with the thesis that intelligence. As we have discussed in the previous chapter. Disambiguation and 9. The chunks are the small-scale structures that people are looking for. A concept and its nearest . As seen before there is the word COMPANY. the essence of the knowledge graph approach is the semantic aspects. A word graph is considered to be without limits essentially. and therefore also artificial intelligence. This too is a typical AI-problem. Although our main goal in this chapter is to show the usefulness of the idea of partial structural parsing in the field of Information Extraction. A Semantic Interpreter. 6. the interpretation as name of a COMPANY is most likely. a COMPANY or a LOCATION. The hardest problems seem to be those encountered in 8. 7. 4. 9. Consider the discussion about Chunk 32. We already saw in Chapter 5 that background knowledge is decisive for obtaining a sentence graph with structural parsing. One of the new features of our approach is that the traditional parse trees get a much less important role to play. Hupplewhite could be the name of a PERSON. We get the filled in templates as knowledge graph structures. 10. In our analysis of the example disambiguation took place by taking into account the other parts of the sentence. A Parser. Disambiguation is context dependent. A Coreference Resolver. Coreference Resolving. A Fragment Combiner. that was solved by taking into account the context. Both are replaced by knowledge graphs. A Lexical Disambiguator. heavily depends on the use of background knowledge. the problems we hit upon deserve some further discussion. A Preparser.180 Chapter 7 of slots could easily be added.

The smaller this lexicon. see Chapter 3. i. that is joined with the first graph.e. For a computer approach. and replacing these concepts by their meaning. i. However. There is. in each of the minds of the speakers A and B. namely the context in which the concept is considered. This becomes even clearer when we consider a dialogue. the fewer the associations the computer has and the less expansion can take place. i. Speaker A says something and a sentence graph is made for this.e. Every time new information is exchanged the graph representing what has been said sofar. The answer of speaker B is likewise transformed into a sentence graph. Like for human beings. the word graph of the original concept can be “expanded” and a larger word graph is obtained. the computer’s abilities to think. however. A person does not have his whole mind graph at his disposal immediately. contained in the lexicon. now coming from the dialogue partner and not from the foregoing text.e. their word graphs. So there are two forms of expansion available to the computer. If. a word graph corresponding to all knowledge available to that mind. the larger these are. for extracting information from the second sentence the computer has the information contained in the first sentence at its disposal too. Next to its internal information. The description of a dialogue by means of knowledge graphs can be as follows. by considering the concepts in the associations. two sentences are given. The subgraph of the mind graph arising after deletion of the concept token can be called background knowledge of that concept. are highly dependent on its information. In principle this can go on indefinitely until the whole mind graph is obtained. there is the external information contained in the context. i. that is simulating this process.e. i. a second source of information. This expansion is also due to context. in the word graph of the concept.3. The more information is contained in the lexicon of word graphs. is expanded. like in our example. In Section 4. the higher the probability that by expansion relevant linking of concepts takes place.Information extraction 181 neighbors form a subgraph of the mind graph that can be called foreground knowledge.1 we pointed out that expansion of concepts plays an important role in thinking.e. link somethings. In a way the context also expands the knowledge of the computer. we have at our disposal the word graph lexicon. One is due to . Given a concept the number of associations with that concept will in first instance be limited.

based on this idea of expansion.182 Chapter 7 combination of word graphs from its lexicon. . is challenging. the other is due to context processing. The development of an automated information extraction procedure.

Columbia. 1984] R. E. 1995. R. MD. 1995] D. Kluwer Academic Publishers. 1995] J. Seventh Edition.. Knowledge Graphs: representation and structuring of scientific knowledge. Ph. Abney. Appelt et al. [Berg. Parsing by chunks. Enschede. [Allen. [Appelt et al. The Pocket OXFORD Dictionary. [Allen. Sixth Message Understanding Conf.. P. In: Proc. Bolc).. In: Natural Language Communication with Computers (Ed. Abney and C. thesis. California. In: Proc. Aberdeen et al. [Bates. Allen.D. Tenny). ISBN 90-9006360-9. Columbia. 1991. MD.. The Benjamin/Cummings Publishing Company. . 1978.. 1995. Oxford University Press. MITRE: Description of the Alembic system used for MUC-6. [Abney. In: Principle-Based Parsing (Eds.. Berlin. The Netherlands. ISBN 90-9001963-4. 1987] J. Berwick. Natural Language Understanding. Inc. Sixth Message Understanding Conf. 1984. 1987] R. 1987. Bakker. University of Twente. L. 1978] M. Morgan Kaufmann. SRI International FASTUS system: MUC-6 test results and analysis. 1991] S. New York. The theory and practice of augmented transition network grammars. Bates.. Knowledge Graphs and Logic: One of Two Kinds. 191-259. R. Enschede.D. van den Berg. 1993] H. Ph. [Bakker. thesis. S. 1987. The Netherlands. Allen. Morgan Kaufmann. Springer Verlag.Bibliography [Aberdeen et al. University of Twente.

Prediction and substantiation: A new approach to natural language processing. Cheng. 1979.. Ph. 1990] G. Foreign Language Teaching and Research Press. Young and G. [Cowie & Wilks. J. [de Vries. In: Handbook of Natural Language Processing (Eds. 1988] C. Ph. Groningen. 1988. Vancouver.D. Medische Besluitvorming: Een Aanzet to Formele Geneeskunde. University of Groningen. Dale. E. On Dou-Quantification. F. de Vries Robbé. MIT Press. Department of Medicine. Cowie et al. Kluwer Academic. 1979] G.D. H. Cowie and Y. Dordrecht. In: Proc. ACM. 1995. 1996] E. [Charniak. 1997] H. [Cowie & Lehnert. Cambridge. Moisl and H. Church. [Church et al. [Ce. Three generative. Collins. de Vries. thesis. 1987] P. The Netherlands. [Blok. In: Special NLP Issue of the Comm (Ed. Lehnert. L. The Diderot information extraction system. [Chierchia & McConnell-Ginet. In: Proc. A Modern Chinese-English Dictionary. 4: 197-234. 1995] L. Information extraction. Somers). Cowie and W. 1997... of the 35th Annual Meeting of the Association for Computational Linguistics. [Cowie et al. Y. [DeJong. In: Proc. Blok. 1993. Chierchia and S. The Netherlands. Corpus-based Methods in Language and Speech. 1989. 1996] J. Cognitive Science. Knowledge Graph Framing and Exploitation. . University of Groningen. 66-23. 1996] K. Bloothcroft. Groningen. Charniak. [Cheng. S. Tree-bank grammars. F. New York: Marcel Dekker. Master’s thesis. 1996. Ce. 1997] M. OR. Representation of Science Texts in Knowledge Graphs. Information Extraction. 1990. of the Thirteenth National Conf. 1031-1036. deJong. 1987. Portlang. R. 2000. University of Twente. 1996. thesis. 2000] J. McConnell-Ginet. on Artificial Intelligence. of the Pacific Association for Computational Linguistics. Wilks). 1993] J. H. The Netherlands. Enschede. 1996.184 Bibliography 1993. lexicalised models for statistical parsing. Wilks. [Collins. Journal of East Asian Linguistics. 1989] P. 3: 251-273. 1997. [de Vries Robbé. Faculty of Mathematical Sciences. Meaning and Grammar: An Introduction to Semantics. of the first Conf.

MD. Grimes.. In: Conceptual Structures: Knowledge Representation as Interlingua. 1968. Ellis and G. of the Fifth Message Understanding Conf. 45. S. J. Li. Grishman and B. Summer Institute of Linguistics Publications in Linguistic and Related Fields. [Harris. Word Graphs: The First Set. Generalized Quantifiers: Linguistic and Logical Approaches. 1993] J. Copenhagen. Harris. [Grimes. Gardenfors. Message Understanding Conference – 6: A brief history. 1995] T. G. [Grishman & Sundheim. Feng and W.. R. Harris. Description of the LaSIE system as used for MUC-6. 1979. 1995] D. MD. 1996. of the Fourth International Conf. F. 1996. [Hoede & Willems. 1982. Virginia. Graph Theory. Bondi Beach. Sydney. on Conceptual Structures (Eds. Mathematical Structures of Language. Gaizauskas et al. Aux. The generic information extraction system.S. [Harris.Bibliography 185 [Fisher et al. Reading. of the 6th Message Understanding Conf. The Netherlands. [Harary. 1987] P. D. no.. Hoede and M. University of Twente. Enschede.. Volume 31 of Studies in Linguistic and Philosophy. W. John Wiley & Sons. [Graham. Sundheim. 1975] J. 1996] R. New York. Australia. P. TAB BOOKS. Fisher. Eklund. Addison-Wesley. Columbia. Knowledge Graphs and Natural Language. [Gardenfors. Artificial Intelligence making machines "think". Reidel Publishing Company. 1972] F. [Harris. In: Proc. 1989. New York. 81-93. A Grammar of English on Mathematical Principles. [Hoede & Li. 1996] C. Willems. D. In: Proc.A. Morgan Kaufmann. S. 1996. Wiley-Interscience. 1995.. U. . Description of the UMass Systems as Used for MUC-6. McCarthy. Harary. 1987. 1985. 1989] C. 1982] Z. Graham. S. 1993. 1979] N. Mann). Network Grammars. 1972. In: Proc. [Hobbs. Morgan Kaufmann. Columbia. 1975. Dordrecht. Proc. 1985] M. Hoede and X. [Gaizauskas et al.. Soderland. Memorandum no. Hobbs. On Computational Linguistics. Harris. In: Proc.811. 1968] Z.. Introduction to Natural Language Processing. Reston. Sixth Message Understanding Conf. A Prentice-Hall Company. Lehnert. 16th International Conf. 87-91.

D. Morgan Kaufmann. 1996] D. Faculty of Mathematical Sciences. In: Conceptual Structures: Extracting and Representing Semantics. Bergen.1453.. Kuno. In: Publication of the Norwegian Computing Center for the Humanities. Hoede and X. Norway. Tools and Applications. ISSN 0169-2690. Liu and L. [Humphreys. In: Proc. 1516. of the 9th International Conf. 1993] W.. Two Applications of information extraction to biological science journal articles: enzyme interactions and protein structures. Lecture Notes in Artificial Intelligence no. Knowledge Graph Analysis of Some Particular Problems in The Semantics of Chinese. In: Conceptual Structures: Theory. Mineau). [Litman. [Krupka. . on Conceptual Structures (Eds.. SRA: description of the SRA system as used for MUC-6. G. Proc. In: Proc. In: Explorations in Cognition (Eds. 1995. Columbia.. Zhang. 1998. [Holland & Johansson. UMass/Hughes: Description of the CIRCUS system used for MUC-5. Freeman. Hoede. 2001. 1975. J. Lehnert et al. X. Montpellier. W. The predictive analyser and a path elimination technique. San Francisco. 1982] K. 505-516. The Netherlands. Delugach and G.186 Bibliography [Hoede & Liu. Cue phrase classification using machine learning. Chein). 375-389. In: Conceptual Structures: Broadening the Base. M-L. 117-135. Word Graphs: The Third Set. Li. 1995] G. CA. Holland and S. On process models for sentence analysis. Enschede. of the 9th International Conf. Morgan Kaufmann. Norman and D. [Hoede & Zhang. Hawaii. 2000] C. Hoede and L. CA. [Kaplan. USA. 2000] Humphreys. 2000. X. 1998] C. 8: 687-698. Sixth Message Understanding Conf. M. USA. 43-53. on Conceptual Structures (Ed. of the Pacific Symposium on Biocomputing (PSB-2000). Kaplan. on Conceptual Structures (Eds. [Hoede et al. 2001. Word Frequencies in British and American English. Rummelhart). Hoede and L. University of Twente. of the 6th International Conf. Litman. Stumme). S. 1982. Mugnier and M. Johansson. Zhang. Proc. 1965] S. In: Proc. Proc. Communications of the ACM. [Hoede & Zhang.. Zhang. Springer Lecture Notes in Artificial Intelligence no. Krupka. Structural Parsing. 1965. 75-88. Liu. A. 2001b] C.2120. H. 1993. E. Aux. Memorandum no. 15-28. [Kuno. [Lehnert et al. Word Graphs: The Second Set. 2000. 2001a] C. fifth Message Understanding Conf. 1975] R. MD.

Cambridge. [MUC-5... Description of the LTG System Used for MUC-7. [Rabiner. 1992] F. thesis. 1997] C. Shabes. 2002. In: Proc. 147-158. Delaware. August 1993. 1965] S.. On the Algebra of Logic. 1997. 1991] Proc. [MUC-6. University of Twente. of the 30th Annual Meeting of the Association for Computational Linguistics. 1993] Proc. 1992] Proc. 1996. Ph. [Pereira & Shabes. [Mikheev et al. Petrick. S. R.C. Inside-outside reestimation from partially bracketed Corpora. of the Fourth Message Understanding Conf. Washington. of the Fifth Message Understanding Conf. 128-135. Columbia. Technical Report CSPR 225. of the Sixth Message Understanding Conf. American Journal of Mathematics. University of Sussex.D.. 1998] A. D. Manning and B. [MUC-3. 1992] C. D. R.. Pereira and Y. Columbia. In: Proc. Morgan Kaufmann.Bibliography 187 Journal of Artificial Intelligence Research... 1992. MD. 1992. Morgan Kaufmann. 1998] Proc. Mass. 1885. The TIC message analyser. Carpenter. Petrick. Liu. Washington. D.D. [Liu. Mellish et al. 1966] S. Moens. Enschede. Thesis. MD. A tutorial on hidden Markov models and selected . Three generative.. [MUC-4. ISBN 90-3651834-2. 2002] X. C. 7: 180-202. A Program for Transformational Syntactic Analysis. Proc. 5: 53-95. MD. Peirce. Columbia. Grover and M. [Manning & Carpenter. MIT Cambridge. 1998.. Wewark. May 1991.. 1995] Proc.. 1885] C. Rabiner. Ph. 1989] L. A Recognition Procedure for Transformational Grammars. The Chemistry of Chinese Language. of the Third Message Understanding Conf. [Petrick. lexicalised models for statistical parsing. of the Fifth International Workshop on Parsing Technologies. Morgan Kaufmann. Morgan Kaufmann. England. 1966. of the Seventh Message Understanding Conf. 1992. Sussex. R. 1965. Mass. [MUC-7. of the seventh Message Understanding Conf. Mikheev. D. [Peirce. [Mellish et al. 1998. [Petrick.C... Airforce Cambridge Research Laboratories Report AFCRL-66-698. The Netherlands. Morgan Kaufmann. 1995.

[Schank & Abelson.S. of the Workshop on Theoretical Issues in Natural Language Processing (Eds. Friedman and M. [Schank. New York. of the IEEE. 1983. Addison Wesley. 1991] H. Sager. In: Advances in Computers (Eds. 1974] C. Scripts. Lyman. J. North-Holland. 1977.. In: Proc. University of Twente. [Riesbeck. F. Enschede.A. C. P. Master's thesis. Abelson. Lawrence Erlbaum Associates.. Plans. 3/4: 552-630. Ph. C. 1975a] C. Yovits and M. NJ. Academic Press. [Sowa. M. New York. Ratnaparkhi. 1977] R. K. [Reidsma. [Radford. McGraw-Hill Inc. L. Cambridge University Press. K. Rich. Artificial Intelligence. Enschede. 1991. 83-156. K. Draft of a book scheduled to the published by the PWS . Schank.. 2001. Nash-Webber). In: Conceptual Information Processing (Ed. 15-20. 1987. 1975] R. Consistency and Robustness of Knowledge Graphs. Medical language processing: computer management of narrative data. Schank). 1974.188 Bibliography applications in speech recognition. Goals and Understanding. Computational Analysis. [Rich.. C. [Smit. Hillsdales. 1975b] C. The Netherlands. The Netherlands. Faculty of Computer Science. 1999] A. Schank and B. [Sager et al. 1975. Schank.D. 1988. Beranek and Newman. Philosophical. Mass. thesis. Riesbeck. Juggling word graphs: A method for modeling the meaning of sentence using extended knowledge graphs. 15:181-237. U. R. Transformational Grammar: A First Course. [Schank. North-Holland. [Riesbeck. Cambridge. 1994] J. [Riesbeck. 1987] N. Bolt. C. 1989. 1988] A. Riesbeck. Cognitive Psychology. 1972. and Computational Foundations. 2001] D. C. Thesis. Cambridge. University of Twente. 1999. Stanford University. Learning to parse natural language acquisition. Knowledge Representation: Logical. C. Schank and R. Reidsma. [Ratnaparkhi. In: Proc. 1972] R. R. Ph. Riesbeck. Department of Computer Science.D. 1975a. 1983] E. Computational Understanding. Radford. 1975b. 77/2: 257-286. Amsterdam. Computational Understanding: Analysis of Sentences and Context. Smit. Sowa. Conceptual Information Processing. Rubinoff). Conceptual Dependency: A Theory of Natural Language Understanding. Amsterdam and American Elsevier.

Rumelhart. A. Willems. Dewar. In: Linguistic Structures Processing (Ed. [Zhu. [Woods. Norman and D. Beijing. Amsterdam. E. of the Seventh Message Understanding Conf. Transition network grammars for natural language analysis. Woods. Commerce Printing House. The Netherlands. [Yangarber & Grishman. Tsinghua University Press. Washington. 1977. P. NYU: Description of the Proteus/PET System as used for MUC-7 ST. 1972. A. Waltz. Edinburgh. 1993. A. L. The syntactic analysis of English by machine. thesis.C. [Stevens & Rumelhart. S..2378. D. 1998] R..D. BBN Report no. Grammar Tutorial. Cambridge. D. 1972] W. 1995. Nash-Webber. Yao. 1995] T. 281-309. [Thorne et al. Enschede. Boston Massachusetts. San Francisco. 1994. An English language question answering system for a large relational data base. 13: 591-606. 1998. X. 1968. Yangarber and R. Thorne. Beranek and Newman. P. M. . Mass. Zhu. 99-11. A. Beijing. North-Holland. E. 1975. Stevens and D. In: Machine Intelligence 3 (Ed.. Communications of the ACM. 1968] J. [Yao. 1978. Woods. Zampolli). [Winograd. L. [Woods. Woods. 1972. 1975] A.. D. Brately and H. [Waltz. [Willems. A. 136155. Understanding Natural Language.Bibliography 189 Publishing Company. 1972] T. 1970] W. Freeman. Bolt. 521-568. Michie). 1984] D. Winograd. Rumelhart). The Lunar Sciences Natural Language Information System. Ph. 1984. 1978] D. 21/7: 526-539. Edinburgh: Edinburgh University Press. In: Proc. Errors in reading: analysis using an augmented transition network model of grammar. University of Twente. Natural Language Processing. [Woods et al. 1977] W. 1970. Also available as publication N72-28984 of the US National Technical Information Service. 1993] M. Edinburgh University Press. Kaplan and B. R. Communications of the ACM. In: Explorations in Cognition (Eds. ISBN 7-302-01911-8.. Chemistry of Language. Grishman. ISBN 90-9005672-6. Lunar rocks in natural English: explorations in natural language question-answering.

190 Bibliography .

Appendix I Word Graphs: The First Set In this appendix a summary is given of the paper of Hoede and Li [Hoede & Li. Although one of the major goals of knowledge graph theory is to deal with linguistic problems. Prepositions form the glue for the more complex words. Before dealing with them systematically we have to discuss a few other things. in particular with pragmatics. The word graph to be included in a lexicon depends on the maker of the lexicon. i. The more associations a mind considers should be made in order to have the proper picture of a concept. The word graph that is actually taken up into a lexicon depends entirely on what the maker of the lexicon wants to express. in which word graphs are discussed for a set of prepositions. nouns and verbs.1 Introduction In building a lexicon of word graphs we should start with the simple subgraphs of the mindgraph. 1996]. the . The meaning of a word is its word graph. For this reason we start the study of prepositions. like is the case with normal lexical. the larger the word graph will be.e. There is no boundary for the word graph. I.

Herewith we have found the first two of our lexicon of word graphs of representations in a very basic form.1 gives the ORD-link and its total graph representation. in this appendix we will focus on minimal word graphs. Whenever words are used in other meanings there should be other word graphs. By this we mean word graphs that express the essence of a word. Like the arcs of the mind graph they glue (some)things together and may be considered the elementary particles of language. given by sixteen different character combinations. in which the basic meaning is recognizable as a subgraph. that gives us the opportunity to introduce the Chinese language.192 Appendix I handling of background information. its most basic meaning. according to our theory. 1978] we read sixteen different meanings for IN. I. As another example. For our purpose we consider the total graph representation of a graph in which arcs are also represented by vertices. Prepositions turn out to be directly linked with the elements of the ontology of knowledge graphs. and type-less arcs represent the structure of the graph.2 Word Graphs for Some Simple Prepositions We consider one simple arc of type ORD. These character combinations express different connotations of the word IN and should therefore be representable by different word graphs. The exercise was carried through to SUB . a: OR b: ORD Figure I. Simple as these graphs are.1(a) has subgraphs like ORD and ORD . Figure I. they express what we would describe by the words FROM and TO. describing those meanings. Its word graph is taken to be . From a Chinese dictionary [Ge. we consider the word IN.

that especially the very simple words will show high similarity between proposed word graphs. OF is a homonym. “attribute of” and “property of”. PAR and FPAR as types of relations that are of merological nature. Note that word graphs depend on the maker(s) of the lexicon.2. where we recognize this triple of types. Figure I. Word graphs for the prepositions OF and WITH. set inclusion. For expressing “part of”-relationships quite a few proposals have been made. Word graph for ZAI4…SHANG4. he/she gives another meaning to the word. The corresponding closest verbal descriptions of these relationships are by “part of”. SUB PAR FPAR SUB PAR FPAR Figure I. essentially.2.3 we discussed logic words of the second kind. SUB ALI ALI area area Figure I. They express. The word graphs of OF and WITH are given in Figure I. The three types are extremely close for the mind that has to bring the relationships “under words”. One of the basic relation types stands .3 shows one of the 16 Chinese words for IN.3. This is a notoriously difficult and much debated field. This is a standard example. In our ontology we have proposed SUB. In Section 4. If a reader does not agree with the proposed word graph. An important preliminary remark to be made is about merology. attribution by the mind and relationship due to framing by the mind. One may expect.Word graphs: the first set 193 find the answer to the question “what is the meaning of this word?” by examining direct translation of the characters.3. The word graph gives the annotation that we are considering areas. however. Here we see a dominant arc of type SUB. The fact that in all three cases the word OF is used shows the source of the discussion in merology.

In the analysis of the Chinese words for IN the SUB-relationship has in several cases not been chosen. we refer to the original paper [Hoede & Li. like we discussed in Section I. For the list of sixteen word graphs. for the word graphs of these words.3 Word Graphs for Other Prepositions In this section the chosen prepositions are mentioned in the grammar of Quirk et al. we do not repeat it here. [Quirk et al. we give the word graph of it in Figure I. Word graph for IN FRONT OF .4. BEFORE ON AFTER ABOVE OVER BELOW UNDER BESIDE BY IN FRONT OF AT BEHIND Take IN FRONT OF as an example. LOCATION ALI SUB ALI PAR ORD ORD ALI ALI ALI SPEAKER PAR LOCATION LOCATION LOCATION Figure I. The following group is used for describing spatial relations mainly. in detail we refer to the original paper. The list of words in this group is: FROM TO OF IN WITH LIKE BY BECAUSE FOR .4. Instead the PAR-relationship seemed to describe the meaning better. The first group consists of the simplest graphs that can be made out of our ontology elements. 1995]. We only list the words that we chose in each group. I. By gradually adding more structure other prepositions arise. because we want to comment on the various groups.2.194 Appendix I central in the word graph.. Here we classify the prepositions into six parts. Ordering in space as seen from the point of view of a speaker can be longitudinal or traversal and in the second case horizontal or vertical. 1972].

· · · · · · · · · BEYOND SINCE UNTIL UPTO ACROSS THROUGH UP DOWN ALONG ≡ ≡ ≡ ≡ ≡ ≡ ≡ ≡ ≡ AT LOCATION PAST AT TIMES AFTER TIME VALUE AT TIMES BEFORE TIME VALUE AT VALUES ORDERED TO VALUE AT LOCATION PAST MOVING AT LOCATION THROUGH LOCATIONS (following vertical ordering) THROUGH LOCATIONS (against vertical ordering) THROUGH LOCATIONS (following path).Word graphs: the first set 195 In the example IN FRONT OF (Chinese: ZAI4…QIAN2 MIAN4. BETWEEN PAST AMIDST AMONGST The remaining prepositions roughly can be divided into three groups. · · · · OFF OUT OF WITHOUT AGAINST ≡ ≡ ≡ ≡ NOT ON NOT IN NOT WITH NOT FOR. which can be recognized in Figure I. A rather large group consists of more complex prepositions in the sense that these can also be expressed in terms of prepositions belonging to the above that we have described. IN and OF have already been described by word graphs. that can be expressed by the NEG-frame. Hence the lack of reference to location etc.4.) we note the tendency in English to break up the graph into smaller parts. literally AT…FRONT FACE. The word graphs of these prepositions are already that large that they can be brought under words by covering the graphs with smaller word graphs. The two other groups involve negation and universal quantification respectively. The prepositions in the next group are often used in more abstract discussions as well. The location of the speaker has been given as reference point. This small list shows the use of negation. .

5. Note that we have chosen not to use the FPAR-relation but its alternative. The word graph for AROUND is essentially AMIDST with focus on locations. BUT FOR (Chinese: CHU2 WAI4. EQU with PAR as word graph. DESPITE. literally OVER MANAGE) and the group WITH REGARD TO. With the group IN SPITE OF. These prepositions have word graphs with strong similarity to the word graph of WITHOUT. APART FROM. Some of these are synonyms. WITH THE EXCEPTION OF. AS TO. The word like BE is so basic that it deserves to be described by the most basic frame. EXCEPTING. These four prepositions show universal quantification for which we can use the SKOloop. BE Figure I. we conclude our discussion of the word graphs for prepositions.5.196 Appendix I There are quite a few prepositions expressing NOT INCLUDING: EXCEPT FOR. WITH ALL (Chinese: JIN3 GUAN3. which is naming a rectangle with the associated word in the corner. EXCEPT. AS FOR (Chinese: GUAN1 YU2….4 Word Graphs for Verbs and Nouns Verbs like BE or EXIST occur frequently in natural language text. FOR ALL. • • • • ALL OVER ALL ALONG ALL AROUND THROUGH OUT ≡ AT ALL LOCATIONS. BUT. . which is the empty frame of Figure I. literally CLOSE AT). WITH REFERENCE TO. I. literally REMOVE OUTSIDE). Word graph for BE.

g. or by means of an attribute POSSIBILITY with value POSSIBLE or an attribute NECESSITY with value NECESSARY.Word graphs: the first set 197 The merological relation types FPAR. in case of transitive verbs. These graphs can be read as “be something with”. by expanding words occurring in the definition. at least. Figure I. HIT ALI CAU CAU CAU SLEEP ALI Figure I.6. in which only the directed ALI-link is used. Word graphs for HIT and SLEEP. HAVE PAR HAVE SUB HAVE FPAR Figure I.4.8. Normal verbs are represented by including two causal links. Word graph for DOG. by replacing the token with a frame. by rectangles with names CAN or MUST. i.6. ALI DOG Figure I. There are.8 is to be read as “something like a dog”. which is described by the word graphs in Figure I. We can expand the concept of DOG to a more complex word graph. The simplest form is the one. see Figure I. e. or one causal link. . In general. Verbs like CAN and MUST can be expressed by POS and NEC frames.e. SUB and PAR are at the basis of the other very basic word HAVE. a word may be related to a large frame.8. in which there is a more complex word graph to explain the meaning of the word. in case of intransitive verbs.7. just like we did for the noun “APPOINTMENT” in Section 7. Word graphs for HAVE. in which a noun may be represented.5. three different forms.

9 is to be read as “something like a dog equal to Pluto”. The directed FPAR-link is used between two tokens. . If a word is the synonym of another word. in principle. like “company and firm” that we mentioned in Section is to be read as “hotdog is a kind of sausage”.8. Here we consider word graphs for nouns in the context of the type hierarchy. For practical applications in terms of knowledge graph theory. like “hotdog”. The directed EQU-link is used between a word and a token to valuate or instantiate the token. Word graph for DOG PLUTO. we will of course need larger word graphs as most of these problems have to be explained by background knowledge and background knowledge is only obtained in larger versions of the word graphs. In their simplest form all nouns and verbs are thus. we give the word graph for it as Figure I. that is very important in a practical NLP systems. COMPANY ALI EQU ALI FIRM Figure I.8. Figure I. Word graph for HOTDOG ISA SAUSAGE. that we mentioned in Section 7.11. described by word graphs. we give the word graph is as in Figure I.4.11.11 is to be read as “company is synonym with firm”.10. EQU PLUTO ALI DOG Figure I. Word graph for COMPANY SYNONYM WITH FIRM. we give the word graph as in Figure I.10. Figure I. such as the IE system. Figure I.198 Appendix I If a word is an instantiation of another word. like “dog Pluto”.9. HOTDOG ALI FPAR ALI SAUSAGE Figure I. The EQUlink is used between two tokens. If a word is a subtype of another word.

It is argued that these three classes of words belong to a general class of words that may be called adwords. For RED we would take the following word graph into the lexicon: EQU RED ALI COLOUR analogous to the graph for DOG PLUTO. 1998].Appendix II Word Graphs: The Second Set In continuation of Appendix I on word graphs for a set of prepositions. adverbs and Chinese classifier words. in this appendix. These words express the fact that certain graphs may be brought into connection with graphs that describe the important classes of nouns and verbs. II.1 Introduction Suppose we have the word combination RED BALL. who studied word graphs for adjectives. The word BALL can be represented. Now RED is a word attributed to the word BALL in the sense that its word graph is linked to the word graph of BALL. by ALI BALL . according to our formalism. Now we say “red is the colour of the ball” . a summary is given of the paper of Hoede and Liu [Hoede & Liu.

Quite a few adjectives are similar to RED and can be seen as adwords linked by the PAR. for example. Suppose.relationship.200 Appendix II and the word graphs are to be linked in a way that we use the word OF for. So the graph might also be brought under words by “(the) ball has colour red”. Yet the adword PI should be expressed in proper speaking. This type of adjective is of another nature than the adjective RED. by means of an EQU-arc. it is more natural to speak of adwords. This is just the view of Hoede and Liu. Also note that we do not say RED COLOUR but do say THE COLOUR RED. One should say “a stick spear”. In Chinese it is not well-spoken to say “a spear”. one time by a PAR-relationship and the other time by an FPAR-relationship. Note that people may differ in opinion on expressing RED BALL. If the type of molecules is denoted in the graph. but lexica contain scores of unprecise definitions). where PI is the classifier. As we see colour as an exterior attribute of ball we choose the PAR-relationship to represent RED BALL by ALI PAR RED EQU ALI COLOUR . It should be clarifying that HAVE was seen as BE WITH. as silicon. that a stone is defined as “a structure of molecules” (which would not be precise enough. A HORSE is not YI MA but should be expressed as YI PI MA. to nouns and to verbs. the difference being expressed by the way of linking. It is usually called an adjective. BALL It is in this way that RED is a word linked to the word BALL. the meaning of which seems to have been lost in the course of time. One particular way is by the FPAR-relationship that links the constituents of a definition to the defined concept. where stick is a word expressing a classifying aspect of spear. BE being represented by the empty frame and WITH having word graph PAR . The essential linguistic phenomenon is that word graphs are linked to other word graphs. As this can be down in various way. we may speak of a SILICON STONE and SILICON is now functioning as an adjective. Note that without the word RED we would still have COLOURED BALL. there are other ways how word graphs can be linked to the word graph of a noun and a verb. There are . We will discuss adjectives and adverbs from this point of view. However. The interesting phenomenon of classifiers in Chinese is closely related to our way of viewing adwords.

in which they give a semantic subclassification of adjectives. i. more specifically its Chapter 5. We will discuss several of them from the same point of view as before and give word graphs for them. we focus on the paragraphs 5. The book of Quirk et al. in a linear order. adverbs and classifiers. but a word graph for BALL that does not mention the possibility of attribution by means of a PAR-relationship. having word graphs that cover the knowledge graph. that enables the short but incomplete utterance path.1 Adjectives As word graphs are supposed to grasp the semantics of a word.Word graphs: the second set 201 more than 400 of these classifiers. The way this is done differs from language to language. there are ways of bringing a knowledge graph under words that are more precise than others.e. As our example RED BALL or BALL WITH COLOUR RED shows. unless it is instructed to interpret the syntactic fact that both words are uttered together as justifying the linking of their word graphs by some arc.2 Adwords In this appendix we cannot give an extensive treatment of the grammatical aspects of adjectives. II.. See also Chapter 6 of this thesis. II.2. A certain knowledge graph is brought under words by expressing by words certain of its subgraphs. ways of uttering the words. For a more elaborate discussion of the interplay of semantics and syntax we refer to Chapter 5 of this thesis. We will not stress syntactic problems. We will restrict ourselves to some major subclasses of these adwords.41 in Quirk et al. 1972] was used for reference. In natural language use often the less precise descriptions.. Any knowledge graph admits an utterance path. are given.37 to 5. A machine may then not be able to create a connected knowledge graph for RED BALL. A lexicon may contain a word graph for RED that includes the colour concept. like RED BALL. Usually there are several utterance paths. [Quirk et al. In an extremely simple example we have that RED BALL in English is uttered as BALLON ROUGE in French. That RED is a colour and that colour can be attributed to a BALL is background knowledge for these concepts RED and BALL. .

They consider BLACK. BRAVER. Before undertaking discussions we should reproduce their premodification examples on page 925 in which combinations of adwords are mentioned. most are gradable and most are inherent. The normal adjective type is all three. determiners etc. One can say BE BRAVE. BLACKER. Their Table 5:2 looks as follows: stative + + + gradable + + + inherent + + + BLACK (coat) BRAVE (man) BRITISH (citizen) NEW (friend) Adjectives are characteristically stative. THE THE A A SOME SMALL INTRICATE OLD general HECTIC EXTRAVAGANT age colour participle provenance noun denomihead nal SOCIAL LIFE LIFE TOWER IDOL DESIGNS LONDON SOCIAL GREY GREEN CRUMBLING CARVED INTERLOCKING CHINESE GOTHIC CHURCH JADE . give the imperative as a way to distinguish stative from dynamic adjectives. BE BRITISH respectively BE NEW. explaining the minus sign in the first column. Inherent adjectives characterize the referent of the noun directly.202 Appendix II They make the distinction stative/dynamic. gradable/non-gradable and inherent/noninherent. Quirk et al. and NEWER are gradings but BRITISHER is not possible. like BLACK. One can say BE CAREFUL but not BE TALL or BE BLACK. explaining the minus sign in the second column. TRUE and BRITISH to be inherent adjectives.

1 The FPAR-adwords We use the FPAR-relationship to represent the definitional contents of a concept. usually some restrictions are made. Braveness is present according to a judgement. For material objects a typical FPAR-adword expresses the sort of material. BRAVE can be seen as an instantiation of judgement. Other adjectives of this type are KIND and UGLY. Another remark that should be made is that a definition may contain the concept of colour. Note that we speak of relative temperature as HOT is not a temperature. BRITISH does not describe an inherent aspect either. If the definition contains a preposition like OF. this word is not considered an adword. Even in white light objects may have different colours for a colour blind person. If STEAM is per definition HOT WATER VAPOUR then HOT is an FPAR-adword as is WATER. for different reason.1.2. as was assumed in the table given. Anything that is part of Britain. In this way the frame name determines the adjective for its constituents. can have with other words. The definition of an elephant may contain the statement that it is a grey animal. or verb. say GREY. call semantic sets. like FAST is . In that case the adjective. For this reason the PAR-relationship is used and RED is considered to be a PAR-adword and colour to be non-inherent in general. Examples of this type of adjectives are LAMB in LAMB MEAT and CITY in CITY COUNCIL or CHURCH in CHURCH TOWER.Word graphs: the second set 203 It is clear that here proposals have been made for what Quirk et al. This is exactly what will be done. however. It is concluded that the restriction inherent/non-inherent had better be replaced by the distinction FPAR/NONFPAR. BRITISH may therefore be called an inverse FPAR-adword. JADE describes the material and is a typical FPAR-adword. instantiating relative temperature and material of vapour respectively. can be called BRITISH. seen as a frame. The point is that colour is attributed subjectively. So in JADE IDOL. is considered to be inherent. Similarly BRAVE and BRITISH are disputable as inherent adjectives. In BLACK COAT. Any word used in the definition might be called an FPAR-adword. but the basis of the proposal will be the types of relationships a noun. II. a red ball in green light looks black. the adjective cannot be considered to be inherent. However. One is considered to be brave by others.

The link to the main concept. when the king was replaced. Adjectives like EARLY and SUDDEN belong to this class and also OLD and NEW.2 The PAR-adwords We use the PAR-relationship to represent exterior attribution. Usually these adwords do not indicate absolute temperature or length but relative temperature and relative length. Within a frame concepts may occur that allow a measure. by exterior attribution. Judgements on a concept are typical exterior attributions. A speech act may occur at time t0. whereas the intended description concerns something at time t1. is by a PAR-link that connects the time aspect to the concept in the following way ALI PAR ALI TIME .2. the adjective NEW refers to a time after time t2. like in the description of tense. BRAVE was already classified as a PARadword.5. BEAUTIFUL is another good example. The adjective FORMER refers to a time before a certain time t2. longer. SOMETHING The word graphs for the adjectives embed the time aspect in a more elaborate word graph in order to express the various meanings. hotter or long. In “a two meter man”. call it. like in hot. the precise length TWO METER may be interpreted as an FPAR-adword too. Its space-time coordinates will be seen as determined from the outside. The fact that this took place before t0 is apparent from the past tense coming forward in WAS. like temperature or length.204 Appendix II not a velocity. i. . In a historical account one may read “the former king was tyrannical but the new king was a very kind person”. a theme that we discussed in Section 6. II. and therefore represented by a PAR-relationship. often determined by the discourse of which the speech act is part. or head as Quirk et al.1. A ball may exist at some location at some moment in time. In those cases the corresponding FPAR-adwords are gradable. An important class of words concerns space and time aspects. In these graphs the time of the speech act may play an important role.e.

Concepts may also be susceptible to subjective measure. which might be part of the frame CHILD. where bad dictionaries give short definitions. with INTELLIGENCE as part of the frame. The more elaborate word graphs may become quite large. In this description. In INTELLIGENT child. if intelligence is seen as something attributed from the outside it is a PAR-adword and we have ALI CHILD PAR PAR ALI INTELLIGENCE ALI MEASURE . and good dictionaries try to give more precise definitions. Definitions may contain concepts susceptible to objective measure. This also holds for OLD or ANCIENT. Differences between the three adjectives must come forward in the word graphs.1. .2. expressing that the first part of the time interval is meant.Word graphs: the second set 205 The word graph for FORMER and NEW will contain an ORD-relationship. Having intelligence. in his choice of words. Depending on the outcome of the measurement the word INTELLIGENT or STUPID may be appropriate. UTTER fool and LOUD noise we have examples of adwords that express subjective measurement. does not justify to speak of an INTELLIGENT CHILD. We mentioned temperature or relative temperature in Section II. INTELLIGENT is an FPAR-adword. In EARLY reference to a time interval will have to be made. However.1 and HOT as an FPAR-adword. This situation is similar for dictionaries. he might speak of “in the early days”. but if these are left out by the speaker. So we have FPAR ALI CHILD ALI INTELLIGENCE PAR ALI MEASURE . “in the old days” or “in the ancient days” to express the same thing. For this some measure should be used which is externally attributed to INTELLIGENCE.

The discussion should have shown that the classification that is proposed depends on the explicit choice of the frames in defining concepts. in decibels.3 The CAU-adwords In the representation of language by knowledge graphs verbs are represented by a token and CAU-relationships. for obvious reasons. BOILING water and MARRIED couple give examples of CAU-adwords. The rightmost token in the first knowledge graph might represent a letter and can be described as “written letter”. Consider the leftmost token. Even if measured. that the word graph should contain information referring to the time aspects. like for NEW. LOUD should be classified as an FPAR-adword. The last example is somewhat tricky. or both are “in the state of marriage”. Transitive verbs like WRITE are represented as WRITE ALI CAU CAU . it is still a subjective choice to say that certain numbers of decibels correspond to LOUD. Then “A and B have got married”. Again we should note. classified as CAU-adwords. Marrying usually involves two persons and “A marries B” and “B marries A”. whereas intransitive verbs like SLEEP are represented as SLEEP ALI CAU . It represents “something writing” respectively “something sleeping”.206 Appendix II Somebody is held to be a fool and any measurement is subjective. Thus WRITING and SLEEPING are adjectives of that something and are.2. So the adjective WRITTEN can also be classified as a CAU-adword. . II. WINNING team. Yet if NOISE is defined by a frame including such a measurement. UTTER refers to the extreme end of an imaginary scale. LOUD is clearly also used for a subjective measurement and therefore classified as a PAR-adword.1.

Suppose something is subject to a series of changes. The check on the distinction stative/dynamic. ASLEEP or ALIVE. that describe states. where descriptions of processes stand central. Having talent or having a disease is expressed by TALENTED respectively DISEASED. focusing on the process rather than on the subject or object related to the process by a CAU-relationship. like in BE CAREFUL. as we stress the relationship with verbs. HAVE. are predicative only.g. which was defined as BE WITH. of escaping respectively retiring. This . Yet the adjectives describing states are discussed in this section. in this case this word IS stems e. but now to express a state after the process. THE ABLAZE HOUSE cannot be said. However. A BOY WITH TALENT is a way of wording that seems better than TALENTED BOY. THE HOUSE IS ABLAZE can be said. Most of the a-adjectives. from A TRUCK IS (DEFINED AS) A HEAVY CAR which is shortcutted to A TRUCK IS HEAVY or even THE CAR IS HEAVY. The analogy is suggested by the word IS. has finished. However. The adjectives in ESCAPED PRISONER or RETIRED EMPLOYEE correspond to intransitive verbs ESCAPE and RETIRE. States are exemplified by adjectives like ABLAZE. CAN and MUST are verbs that likewise correspond to the frame relationships that were distinguished. It should be noted that predicative use of an adjective. THE PRISONER WHO IS ESCAPED respectively THE EMPLOYEE WHO IS RETIRED seem better ways of wording than THE ESCAPED PRISONER respectively THE RETIRED EMPLOYEE. they usually express states. Strictly speaking the two adjectives are not CAU-adwords. In THE CAR IS HEAVY the use of the adjective HEAVY is predicative. corresponds to checking whether the adjective can be seen as describing a state. In Turkish DOKTORDADIR literally means DOCTOR AT BE. by trying out the imperative. The form suggests a verb. expresses a BE-frame. In HAPPY girl the adjective describes a state that the noun is in. THE CAR IS HEAVY instead of THE HEAVY CAR expresses the “being heavy” of the car explicitly. Like for objects of a transitive verb the –ED ending is used. The very frequently used agglutination DIR expresses the BE-frame.Word graphs: the second set 207 This brings us to a special class of adjectives. When they are used. like ABLAZE. “girl being happy” is doing the same. At any time it is then in a certain state. The verb BE was represented by the empty frame.

nouns acting as adjectives. Prototype definitions express what has been seen as common properties of a set of somethings. This means that there is no basic difference in the way other words may act as adwords of verbs in comparison to the way other words act as adwords of nouns. for example in combination with BEHAVIOUR. II.208 Appendix II explains the predicative use of an adjective that is essentially an FPAR-adword.2.4 The ALI-adwords The ALI-relationship between two concepts expresses the alikeness of the concepts. This relationship may be seen as primus inter pares as the process of concept creation seems to depend heavily on the becoming aware of alikeness. DISASTROUS expresses a similarity with a disaster and INDUSTRIOUS expresses a similarity with certain aspects of industry. Other endings are –ISH as in FOOLISH or –LIKE as in CHILDLIKE. Time and location of the act expressed by the verb are natural aspects to consider in .2 Adverbs Nouns and verbs are describing concepts basically in the same way. HEAVY is pushed into the role of a state describer like ABLAZE. words of the second type do. Whereas words of the first type do not necessarily include CAU-relationships in their definition. But for this difference a verb may be seen as a special type of noun. This is also underlined by the possibility of substantiation of verbs. respectively. Adjectives that are expressing that a concept looks like another concept may have specific endings. it becomes clear that the classification that we give in different types of adwords makes sense. A special category of adjectives is ALI-adwords that are themselves nouns. We will discuss only a few examples. Especially for this category of adjectives. THUNDER in THUNDER NOISE or TRAITOR in TRAITOR KNIGHT expresses that the noise is like heard when thunder occurs or that the knight acts like a traitor. PLAYING and A PLAY.2. Compare TO PLAY.1. II.

in most case. An important aspect of this example is that the sheer possibility of compiling adwords in language is an argument for the modeling of language in terms of knowledge graphs. EXTREMELY. HOWEVER. MUCH. ALI-adwords include CLOCKWISE. . We mention OFTEN. The great advantage of our theory is that this discussion is avoided. showing compilation of adwords. OUTSIDE. NEVERTHELESS. Instead one may have the discussion about the classification that is be given to each of the adwords. there is a set of words. WELL. Hence adverbs like AMAZINGLY or SURPRISINGLY may be mentioned. built from word graphs. hence the treatment of these two aspects at the same time. Hoede and Liu prefer to include them in a special third list of words graphs with NO and PROBABLY. QUITE. the way these word graphs glue together. that are sometimes used as adverbs. Firstly. but actually refer to the use of logic in language. In traditional discussion we would have to decide whether we are dealing with adjectives or adverbs. FPAR-adwords are somewhat rare. is made out of the word graph. but also words like TOO or AS are used as adverbs and clearly should be another subclass of these adwords. Secondly. ENOUGH. or any of the many adverbs with ending –WISE. that expresses the text. A FAR MORE EASILY INTELLIGIBLE EXPLANATION. EVER as examples.. CAU-adwords refer to influences or. Chapter 4 of this thesis focuses on such logic words. to consequences of the acts described by the verbs. far more so than for nouns. mentioned by Quirk et al.Word graphs: the second set 209 relation to the verb. The country’s economy must be seen as frame part and for this reason ECONOMICALLY may be classified as an FPAR-adword. SO. ALMOST are reflecting judgements or measurements. Two remarks are to be made still. once a knowledge graph.. A judgement may be interpreted as a subjective measurement. The explicit word graphs for these words may become quite large as rather complex aspects are expressed. immediately gives the answer in that discussion. by which type of arc. ELSE are such words. BRIEFLY. YET. to mention some other potential adverbs. we would like to comment on an example of Quirk et al. Judgements and measurements are two aspects that are also often expressed by PAR-adwords. Many adverbs refer to these aspects and are classified as PAR-adwords. However. THOUGH. In “the country is deteriorating economically” the adverb ECONOMICALLY indicates in what respect the country deteriorates.

1 FPAR-classifiers One of the most frequently used quantity words is GE. Another example YI FENG XIN for A LETTER.210 Appendix II II.2. ALI FPAR NOUN ALI QUANTITY WORD . it should refer to something felt as an essential property of the following noun.3 Classifiers in the Chinese Language In Chinese. the hole property of a well is clearly mentioned as property not to be deleted in the description. Quantity words like these should perhaps better be called property words or classifiers. Here. a special class of words is formed by the quantity words. as Chinese prefer to call them. YI GE DONG XI. THING. For A WELL we have YI YAN JING. II. used in combination with a word like DONG XI. Many quantity words express the property of the noun that it describes a unit. There are many nouns for which the quantity word GE is used. However. Although a word like GE is not felt to have a meaning of its own by Chinese. there is a verb FENG for folding and it is clear that XIN (LETTER) is described in a way that expresses the property that letters usually consist of folded paper. that in the combination is not felt to have a meaning.2. the quantity word is an FPAR-adword and. where the quantity word YAN has some meaning. which are also called classifiers. namely HOLE. in terms of knowledge graphs. For A THING in Chinese we should say. Hence here the relationship between noun and quantity word is an FPAR-relationship.3. indicates the something described by the quantity word. . shows an adword FENG. As this example already shows the quantity word may be seen as a word naming a subframe of a frame carrying the name of the noun. we get the following graph. If a well is the configuration of water coming from a hole in the ground.

2. In knowledge graph representation we have in first instance LIQUID ALI WATER EQU PAR ALI PAR ALI POT ALI SPACE SUB SPACE .3. but are used in combination with a noun to express a certain feature. but the containment feature gives the direct link to the noun. HU. in fact our example is a clear example of a relationship in knowledge graph theory. So HU is an adword of SHUI linked to it by a relationship of complex nature. The important feature is that of containing a liquid. The graph can be extended to include the other features of POT like shape or material. In terms of adwords we cannot say that POT is a PAR-adword. the linking to WATER is more complex. there are many other classifiers that do have a separate meaning. . metal or class. as a container it has a specific shape and is made of some material like China. There are quite a few adwords of the type CONTAINER. is an adword for SHUI. We will not discuss this in more detail now and only mention a typical example: YI HU SHUI. In English too a class description of water is necessary.2 Other classifiers Next to the FPAR-relationship to a noun of a property word. or POT. A POT OF WATER. that is however representable by the basic types of arcs. For the list of word graphs for the other 12 classifiers.Word graphs: the second set 211 II. If we have to baptize this relationship we would choose the word CONTAINER or CONTAINMENT. or WATER. see the original paper [Hoede & Liu. A pot typically contains a liquid. These quantity words can be divided into 13 subclasses. Two concepts are part of one graph and it is this graph that characterizes the relationship between the two concepts. PING or BOTTLE is just one other example. 1998].

212 Appendix II .

106 attribute 24 causal relation 22 chunk IX chunk indicators 4 compositionality principle 34 computational linguistics 2 concept 3. 201 argument transition network 9 artificial intelligence 1.Index adwordial phrase 84 adwords 84. 19 concept identification 36 concept integration 37 conceptual dependency 16 context 181 coreference resolution 145 equal 23 event 143 existential graph 29 functional 25 frame 26 general 43 grammar 7 IE 144 informational dependency 25 inherent 51 integrated 16 joining sentences 3 KG-definition 154 KG-extraction 150 knowledge graphs VII knowledge representation 1 dependency relation 25 distribution 132 . 130.

17. 108 value 23 word types 72 .214 Index logic words 4 logic words of the first kind 45 logic words of the second kind 48 mappings 25 mind graph VIII. 69 syntactic sentence graphs 17 syntactic word graphs IX. 17. 71 sentence graph VIII structural parsing IX. 4. 17. 71 template 145 token 20 totality 124 transformational grammar 9 type hierarchy 21 u-arc 139 utterance paths IX. 4. 28 MUCs 143 MT 144 Name Entity recognition 145 natural language processing 2 natural language understanding 2 NECPAR 26 NEGPAR 26 nonmovable adverbs 131 non-syntactic 16 ontology 27 ordering 25 other logic words 44 parsing 7 partial structural parsing 151 phrase marker 110 POSPAR 26 property sets 24 pure logic words 44 relation integration 37 relationship 19 scenario 145 semantic chunks 17 semantic sentence graph 17 semantic word graphs IX. 17.

because it can replace the others. Section 1. Chapter 3 focuses on the knowledge graph ontology and makes comparisons with a few other well-known ontologies.Summary In Chapter 1. till now no one can be called universal. we may say that O1 is superior to O2. As knowledge is represented in terms of language. Given two ontologies O1 and O2. Each word not only has specific meanings. A comparison with the several hundreds of ontologies known poses a challenging problem. an outline of the contents of this thesis is given. This is why the knowledge graph theory was put forward and has been developed gradually. in the knowledge graph project the focus . its semantic aspects. ● Although there are many ontologies for knowledge representation. In this summary we focus on the background of the problems considered and on future research items. This thesis mainly addresses the following points. We propose how to express words by word graphs both semantically and syntactically. in which a word is a basic component. Chapter 2 gives a short account of parsing natural language for those readers that are not familiar with that subject. if the elements of O2 can be expressed in terms of the elements of O1 and not the other way around. the first goal must be to express language in terms of the knowledge graph ontology. However.3. ● Human language is a kind of special symbol system. its syntactic aspects. but also has a structural function as it plays a role with respect to other words.

i. both for English and Chinese has been published. The third set consists of the group of so-called logic words and is described in Chapter 4. As one of the goals is to develop translation systems. ● Applications of knowledge graph theory are a challenge. by hand. the second set consisted of adjectives. The development of computer programs for uttering a sentence graph.e. A point of further research is to see whether some more sets are to be included for English and Chinese. This in particular with respect to the artificial intelligence aspects that have come forward in our explicit treatment. which aims at transforming an input sentence into a sentence graph. can be uttered. we developed a method for carrying out Information Extraction (IE). parsing is more special than traditional parsing and is called structural parsing. verbs and nouns. and to see whether some special word classes exist in other languages.e. transformation of the sentence graph and uttering the sentence graph in the target language.216 Summary first was on the semantic word graphs. The first set of words consisted of prepositions. giving a meaning to this sentence. with the main steps: structural parsing. the implementation of this method poses an interesting challenge. For mapping a sentence on a sentence graph. The sentences arising from these ways of uttering consist of words occurring in the sentence graph in a specific order. A series of papers on making word graphs. . both syntactic and semantic information is needed. We argue that the structural parsing developed in Chapter 5 of this thesis could be used both in English and in Chinese. as well as for the transformation of sentence graphs from one language into another. i. Languages differ in the way the words occurring in the sentence are ordered. of an example. especially in NLP. ● Given a sentence graph there are usually several ways how such a graph can be brought under words. This is studied in Chapter 6. adverbs and Chinese classifiers. The appendices I and II give a survey of the first two sets mentioned. that forms the contents of Chapter 7. In the context of knowledge graph theory. Based on the theory developed in this thesis. ● Parsing is an essential part of Natural Language Processing (NLP). should also be part of a project on automatic translation based on knowledge graph theory. Again. an important point of further research is to develop computer programs for structural parsing.

From September 1983 to July 1987. she was appointed to study Artificial Intelligence under the supervision of Prof. In 1999. Zhou Guodong. Dr. On May 1996 she won the first prize of the “Teaching Competition” held by the same university.Curriculum Vitae The author of this thesis was born on March 10. . she worked as a teacher at the Department of Computer Science of the Northwest University on July 1990. student under the supervision of Prof. In July 1987. 1964 in Xian. Dr. From 1971 until 1983 she attended primary and grammar school in her home town. she was appointed associate professor. Cornelis Hoede from the University of Twente and Prof. In the course of her studies she concentrated on knowledge representation. After obtaining the master’s degree. Shaanxi Province in the northwest of China. and her master thesis was titled “An Expert System Tool Based on Conceptual Structure”.D. Li Xueliang from the Northwestern Polytechnical University. Here the author investigated conceptual graph theory and its applications in building an expert system. In January 1998 she started as a Ph. she studied computer software at the Northwest University.

Sign up to vote on this title
UsefulNot useful